AI4Science · Daily Brief

AI4Science Daily · 2026-10-13 — This is normal bad work

Tue, 13 Oct 2026 11:12:03 +0000

Selected developments

This is normal bad work
[This is a guest post by Sophie Huiberts , being an expanded version of a post on her own blog . This blog post was initially written in a different file format and converted using AI. — T.] When you prove a theorem, you are imparted a responsibility to the field of mathematics. At minimum you need to explain what you did, including which steps required new ideas and which steps are already well-known. More substantively, you are expected to nurture the literature. If an important idea was poorly explained in its first incarnation, then you have a duty to explain it better in your work. The responsibility is yours because you have the opportunity to publish a paper on the subject. These demands are broadly accepted and are neither new nor controversial. Any time you exceed the demands, the resulting paper is highly appreciated and valued. Any time you fail the demands, you will get a “revise and resubmit” for a journal submission, or a straight rejection for a conference. Until you fix your exposition, your ideas will not enter the canon. If you post unpolished manuscripts to the arXiv, you are embarrassing yourself. Compare this to OpenAI’s behavior. They do not care about their “preprints”. They don’t bother putting in appropriate attribution of ideas, they salami-slice 372 results over 719 pdfs, their work contains retraction-worthy errors , they promise code but neglect to include it , and their proofs are fiendishly hard to parse. Their pdfs actively set back the state of the literature. This is poor quality work and these pdfs are all easy rejects. We all know that OpenAI is not doing this “for the good of science”. These pdfs are exhaust fumes from their internal benchmarking effort. We do not need them to release this unprocessed exhaust into the atmosphere. OpenAI’s new internal model may have produced these texts with the press of a button. But similar results would have been found soon by experts, with or without LLM prompting. Some results were already found by ChatGPT users and were in the process of being written up properly. Even without alleging theft, it is a bad look for a supplier of scientific software to actively scoop their customers. If an expert prompts a result, they take up the mantle of responsibility I describe above. OpenAI not only refuses to take their responsibility, but they prevent their own customers from doing so. October…
Terence Tao · Original source

AI4Science Daily · 2026-10-12 — Import AI 476: Humanity’s Sixth Sense benchmark shows AI visual reasoning lags humans

Tue, 13 Oct 2026 04:38:27 +0000

Selected developments

Import AI 476: Humanity’s Sixth Sense benchmark shows AI visual reasoning lags humans
Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe now New benchmark shows how AI models lack intuitive visual reasoning: …Humanity’s Sixth Sense highlights some shortcomings of current AI systems… AI startups Elorian and Scale AI have built and released Humanity’s Sixth Sense, a benchmark that “measures how a model interprets and reasons through information a visual scene implies but doesn’t explicitly show.” Some examples of the types of task in HSS are a picture of a bookcase and being asked whether it can fit more books, or a photo of horses pulling a plough and a question of whether the chains bear the plough’s drag weight. HSS contains 522 questions that cover 288 images and 234 short video clips (totalling 17.6 hours of video). The questions fall into four categories: Temporal & causal dynamics: “What happened before the scene, and what happens next.” Physical & spatial logic: “What fits, what’s in reach, and what someone else can see from where they are standing.” Social understanding : “Intent, attention, belonging.” Abstract & contextual inference : “The patterns and rules a scene implies.” Results: The human baseline for this is not perfect, but is very high, with people scoring aggregate pass@1 of 93.1, broken down across four categories: Image (94.1), video (91.9), time (97), and space (94.9). By comparison, the best scoring AI system, GPT 6 Astra, gets pass@1 of 53.6, across image (55.6), video (51.1), time (58.2), and space (53.6). Interestingly, Chinese open weight models do poorly here, with GLM-5.3-flash getting a pass@1 of 26.8, and Kimi K3 getting pass@1 of 25.5. “The two most common single mistakes were (1) missing a detail that decides the answer (21% of failures) and (2) misidentifying an object, person, or role (20%),” they write. Why this matters - HSS tells us whether world models are emergent: One way of looking at HSS is highlighting deficiencies in current systems, but a more interesting way is highlighting how adept today’s frontier LLMs are at doing tasks which many people have long assumed would require sophisticated, different AI systems to solve. For instance, many people have claimed that true visual reasoning will rely on developing spatially aware 2D and 3D world models, then using that for reaso…
Import AI · Original source
Generative AI has changed mathematics forever. Where to from here?
By Melissa Lee , Monash University In August, artificial intelligence company OpenAI announced ten advances in mathematics and computer science made with their as-yet unreleased model Astra. The discoveries cover a wide range of mathematical fields, including geometry, cryptography and coding theory. These ten are just the latest in a rapidly growing list of mathematical breakthroughs made by generative artificial intelligence (AI) systems. OpenAI accompanied the announcement with a statement on “responsibility to the mathematical community”. Here, the company acknowledges concerns about attribution, accountability for the correctness of AI-assisted results, and the changing nature of mathematical discovery. The statement reflects a growing understanding that the remarkable technical advances due to AI are only part of a bigger story about what mathematics means, and what its future will look like. AI is raising big questions Large language models (LLMs) such as ChatGPT and Claude are creating disruptions in all parts of the mathematical community. University departments are grappling with the ethical implications of these tools in their research and teaching. Editorial boards of many peer-reviewed journals are being inundated with AI-written papers and must decide how to assess and disclose the use of generative AI. The arXiv preprint repository has seen a sharp increase in mathematical submissions in recent months. Math papers uploaded to @arxiv per month. Jan 1992 to Jul 2026. Source: https://t.co/nrqjhJBtij pic.twitter.com/pmHvtbYK2X — Pablo Groisman (@pgroisma) August 3, 2026 Funding agencies are starting to develop policies governing acceptable use of AI in applications and assessment. Broader philosophical questions about how and why mathematics is done in the first place are hard to avoid. If an LLM can discover a proof, construct an example, or formulate a new question, is there anything uniquely valuable about human mathematical creativity? Is mathematics mostly about producing new theorems, or is it about developing understanding? When a machine contributes to a discovery, who should get the credit? Growing concern – but little consensus Among working mathematicians, there is little consensus about the answers. Some have genuine anxiety about the future of the discipline. One mathematician wrote that the advance of AI-assisted mathematics had tri…
AIhub · Research news and interviews · Original source

AI4Science Daily · 2026-10-10 — Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub

Sat, 10 Oct 2026 10:19:22 +0000

Selected developments

Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub
From the Bitter Lesson of AI scaling to the unsolved mysteries of protein folding , Google DeepMind’s Pushmeet Kohli and Biohub’s Sal Candido are rethinking what it takes to build AI that truly understands biology. In this special panel moderated by Brandon Anderson , they explore why AlphaFold’s breakthrough was only the beginning , why scaling compute and data alone won’t solve biology, and how the next generation of AI models could transform our understanding of proteins, cells, and human disease. We go deep on the future of AI-driven biology : finding scaling laws in biological data, the tradeoffs between scientific intuition and general-purpose architectures, why protein structure prediction is far from solved, and what it would take to build predictive models of living systems. Pushmeet reflects on the lessons behind AlphaFold , the limits of human interpretability, and whether future frontier models could understand other AI systems better than we can. Sal explains why protein language models may already contain scientific knowledge we haven’t unlocked, how biological modeling must move beyond individual proteins, and why achieving Biohub’s mission to cure all disease requires thinking in terms of 10x breakthroughs rather than incremental improvements. We discuss: The Bitter Lesson for biology : why scaling compute and data isn’t enough Why finding the right scaling law matters more than blindly increasing model size How low-quality metagenomic data can improve protein language models Why AI researchers optimize for available data instead of the most important scientific problems Lessons from DeepMind on balancing modeling, data generation, and scientific expertise Why building a virtual cell requires fundamentally different datasets AlphaFold’s handcrafted architecture and the role of scientific intuition Why good data matters more than simply having more data Inductive biases, scaling laws, and the future of specialized AI architectures Why we aren’t in a post-Transformer world , but architectures are evolving Why AlphaFold didn’t actually solve all of protein folding Protein dynamics, disorder , and the limitations of static structure prediction How cryo-EM micrographs could unlock richer biological representations Moving from models of individual proteins to whole biological systems Feynman’s famous principle and why AI can now create things we d…
Latent Space · Original source

AI4Science Daily · 2026-10-09 — The missing map of the sky: the first complete ultraviolet view

Mon, 12 Oct 2026 00:50:42 +0000

Selected developments

The missing map of the sky: the first complete ultraviolet view
Astronomers have mapped the entire sky in visible light, infrared, radio, X-rays, and gamma rays. No one, however, had created a complete map of the sky in ultraviolet (UV) light. Here, Brice Ménard, an astrophysicist at Johns Hopkins University and a researcher at Anthropic, explains how he worked with Claude Science to produce the first complete map of the sky in UV light.
Anthropic · Science Blog · Original source
Periodic Labs on "Synthesis Superintelligence": Training AI Scientists on Real Experiments
It’s hard to believe that Periodic was only launched last September : One year later, it is considered one of the pre-eminent AI scientist labs, with dizzying talent density and astonishing progress in the autonomous lab buildout: Most people are familiar with the standard credentials of Liam and Dogus , but we found an incredible “talent slope” while learning more about Periodic, where each successive employee seems more impressive than the last: From building AI systems that reason over noisy physical experiments to creating laboratories where every instrument can become intelligent, Periodic Labs is betting that the next frontier of AI won’t come from simply training on more internet data , it will come from letting models experiment with the real world. In this episode, Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk join swyx and Brandon to explain why scientific discovery is fundamentally different from math and coding, and what it takes to build AI scientists that can actually discover new materials. We go deep on Periodic’s vision for “synthesis superintelligence” : reinforcement learning grounded in physical experiments, AI-powered materials characterization, simulations and density functional theory, high-throughput labs, and systems that learn from the entire process of doing science rather than only its published results. Liam and Dogus also explain why frontier models will still need experiments , why failed experiments may be some of the most valuable training data, what it means to give every piece of lab equipment “140 IQ,” and how autonomous experimentation could compress decades of scientific trial-and-error into months. We discuss: Why intelligence alone isn’t enough for scientific discovery How reinforcement learning changes when the environment is the physical world Why science requires reasoning under uncertainty, noise, and missing information Prediction, synthesis, and characterization in the materials discovery loop Why physics and materials science are still far from “solved” The “matter compiler” and Periodic’s goal of synthesis superintelligence Phase transitions, X-ray diffraction, and AI-powered materials characterization DFT, simulations, and why experiments remain the ultimate ground truth Room-temperature superconductors, new magnets, batteries, and more efficient compute Why quantum computing may not automatically solve mate…
Latent Space · Original source
Ai2 at COLM 2026: Open research, from models to agents for science
Ai2 at COLM 2026: Open research, from models to agents for science October 9, 2026 Ai2 This week Ai2 was at COLM 2026 in San Francisco. The work we shared reflects our wider mission: building AI research artifacts anyone can study and extend, and developing AI for science tools in collaboration with domain experts. Our evening gathering gave colleagues from across the COLM and wider open-source community another setting to connect. Those exchanges are central to how we approach research. As a fully nonprofit lab, we pursue fundamental scientific questions while sharing the models, data, code, and methods that let others investigate them too. COLM brings us together with the people who use our resources we release, challenge our assumptions, and contribute their own ideas. Sharing the work in open collaboration with others isn’t an afterthought for us—it’s a fundamental part of how we approach research. From Olmo Hybrid to the next Olmo Our paper “Olmo Hybrid: From Theory to Practice and Back”, presented at the conference, examines what happens when a language model combines transformer attention with linear recurrent layers. Attention can retrieve specific details from earlier text; recurrent layers maintain a compact state that updates as the model processes a sequence of text tokens. The research connects theory about what these architectures can represent with controlled training experiments. Olmo Hybrid provides evidence that combining attention and recurrence can improve training efficiency: in the study, it reached the same accuracy as Olmo 3 7B on MMLU, a benchmark covering knowledge across many subjects, using 49% fewer training tokens. That research is helping inform the next Olmo model, now in pre-training with a hybrid mixture-of-experts architecture. The hybrid design combines attention and recurrence. Mixture-of-experts offers another way to improve efficiency, routing each token through a subset of the model’s components. We are also opening up the infrastructure behind this direction. Released ahead of COLM, Olmo-core 3 adds a redesigned system for training large mixture-of-experts models. Researchers can use the code to develop their own models and investigate training decisions, alongside the checkpoints and technical reports available through the broader Olmo ecosystem. And Google recently announced that they recreated our Olmo 3 7B traini…
Allen Institute for AI · AI for Science · Original source
Advancing chemical reaction prediction in data-scarce drug discovery with active and geometric deep learning
Nature Computational Science, Published online: 09 October 2026; doi:10.1038/s43588-026-01063-0 Researchers pair active learning with geometric deep learning to predict the outcome and site selectivity of chemical reactions from scarce data, guiding experiments that accelerate drug discovery and expand accessible chemical space.
Nature Portfolio · Nature Computational Science · Original source
Dataset of DInSAR wrapped phase signals for AI-based automated detection and classification of mass movements
Abstract With the growing use of Artificial Intelligence (AI) in remote sensing of mass movements, open and suitable datasets for model training and validation are increasingly needed. Although Differential Synthetic Aperture Radar Interferometry (DInSAR) is a widely used technique for studying mass movements, wrapped interferograms remain less exploited, and the importance of geomorphological expertise in their interpretation is not usually emphasised. In this work, we introduce a dataset of DInSAR wrapped phase signals designed to support the development of Deep Learning (DL) models for the automated detection and classification of active slow-moving mass movements. The dataset covers two selected wide areas in the Central European Alps and the Northern Apennines. It contains 4910 expert-annotated DInSAR wrapped phase signals derived from 92 Sentinel-1 interferograms with temporal baselines ranging from 6 days to 1 year, classified into nine distinct landslide and periglacial landform classes after careful geomorphological interpretation. This dataset is expected to support the scientific community in AI-based applications for mass movement research, while also serving as a benchmark for the generation of comparable datasets.
Nature Portfolio · Scientific Data · Original source

AI4Science Daily · 2026-10-08 — Complementary remarks from Gary Marcus and Terence Tao on OpenAI's giant math drop - Marcus on AI | Substack

Sun, 11 Oct 2026 10:14:24 +0000

Selected developments

Complementary remarks from Gary Marcus and Terence Tao on OpenAI's giant math drop - Marcus on AI | Substack
OpenAI’s massive new math drop: [This essay was written in extreme haste before a very long wifi-less flight; please forgive typos.] Part One: My take The real news here isn’t the result; it’s not what we were told. 1. AI once tried to be a science. Now we get stuff like the completely vague report from OpenAI below: “Same procedure”? “Using an unreleased model”? This would never pass peer review. We don’t know what the procedure was. We know nothing about the architecture. For esxample, were the proofs generated in one shot, and then verified by the symbolic system Lean? Was there an iterative process?) We know nothing about the failure rate. We know nothing about the training/post training/data augmention. 2. As a result, we have zero idea of how generalizable the result is outside math. 3. A lot of the discussion on social media has been reduced to an ignorant cheering section that applauds without knowing what it is applauding or what it might mean— without ever asking basic scientific questions. The new system could be a legitimate step toward AGI. Or it could just be a clever leveraging of Lean and synthetic data in a verifiable domain with no generality whatsoever. From the initial report, we can tell almost nothing. §
Discovery · OpenAI · Original source
China's XtalPi debuts KodexiaTM as world's first closed-loop siRNA platform
image credit- shutterstock China's XtalPi Holdings, an AI and robotics-driven drug and materials discovery company, has debuted Kodexia™, the industry's first closed-loop small interfering RNA (siRNA) discovery platform combining generative AI, first-principles-driven biological modeling, and automated experimentation. Validating the rapid execution capabilities of this new infrastructure, the company simultaneously announced the advancement of six proprietary therapeutic programs, marking the successful expansion of XtalPi's technology into nucleic acid therapeutics. Kodexia™ also globally co-optimises dual-target siRNAs as single molecules, creating a strategy for disease biology that may not be adequately addressed through single-target intervention. The platform's lead programme in immunoglobulin A (IgA) nephropathy serves as an early empirical validation of this unified architecture. programme is on track for preclinical candidate selection within a nine-month project timeline, significantly accelerating the industry norm of 12 to 18 months. Kodexia™ currently drives six proprietary programmes across metabolic, renal, respiratory, and central nervous system diseases, with more than half having already completed in vivo efficacy evaluations. - China's InnoCare inks $3.5 B deal w...28 September 2026 - Novo Nordisk receives Singapore app...24 September 2026 - Ansell injects $60 M in India to es...21 September 2026 - Avantor expands RIM single-use port...17 September 2026
Discovery · XtalPi · Original source
Researchers fear being scooped by AI, as OpenAI's announcement timing draws questions
As AI agents’ ability to do science autonomously continues to improve, scientists are facing a once-unthinkable question: will I get scooped by a bot? At least twice in the past five weeks, researchers have said that they had been working for an extended period on a particular research question, only to learn that artificial-intelligence companies had either answered the question before them or announced results before they were ready to do the same. In response, researchers who are worried that AI tools might scrape their unpublished work and scoop them are limiting their use of such tools. Anthropic’s AI biolab finds ‘CRISPR-like’ DNA in viruses. What’s next? Nevertheless, some AI researchers, although not all, say that AI agents are increasingly likely to reach conclusions before scientists can do so. Researchers are “going to be scooped, but not because they’ve uploaded a manuscript”, says Geoffrey Irving, chief scientist at Resolution, an AI-safety research organization in Berkeley, California. “They’re going to get scooped because the AIs are very good at solving problems, and they’re going to get better and better.” Controversial findings On 7 September, Tristan Buckmaster, a mathematician at New York University in New York City, posted online that he and a collaborator had made progress on the Navier–Stokes problem, a long-standing open question in fluid dynamics. According to Buckmaster’s account, he and Levent Alpöge, a mathematician at the AI company Anthropic, based in San Francisco, California, whose work with Buckmaster was a personal project, had been pursuing the problem for a year. Buckmaster’s 7 September post also included a paper with the solution to a simpler version of the problem. The next day, 8 September, OpenAI, also in San Francisco, announced that its agents had solved the puzzle — a timing that Buckmaster has publicly questioned. Who gets credit in the AI era? OpenAI maths bombshell sparks debate Buckmaster raised the possibility that information that he and Alpöge had uploaded to an OpenAI tool could have been used to train the company’s models. OpenAI disputed this scenario. The company said in a statement that an investigation had confirmed that Buckmaster’s “prompts” in the two months before the 8 September announcement “could not have influenced the system in any way, including through training”. The statement also said tha…
Nature Portfolio · Nature · Original source
A Closer Look at Agentic BBO: Benchmarking LLM Agents for Black-Box Optimization
Black-box optimization (BBO) arises in many scientific and engineering problems where objective evaluations are expensive and limited. Recent large language model (LLM) agents offer a new way to approach BBO by combining task semantics, computation, optimization tools, and feedback-driven decision making, showing great potential due to the integration with mathematically rigorous tools. However, existing agentic BBO studies use different task domains and system configurations, making their results difficult to compare and the effects of individual design choices hard to isolate. We therefore introduce AgenticBBO-Bench, a cross-domain benchmark for agentic BBO spanning synthetic functions, hyperparameter optimization, database tuning, chip design, and molecular design under a unified finite-budget evaluation protocol. In our experiments, agentic BBO achieves higher family-averaged scores than direct LLM-based methods in all five domains and outperforms the best numerical optimizers in four. We further study three factors shaping agent performance: optimization tools, task information and prior knowledge, and the role of the LLM during search. Our results show that additional numerical tools do not consistently improve performance, task semantics are broadly useful while more specific priors are less reliable, and numerical optimizers can effectively absorb gains from search trajectories established by the agent. Finally, we introduce a five-task frontier challenge within AgenticBBO-Bench and evaluate seven LLMs under the Codex agent harness, where GPT-6 Astra and DeepSeek-V4.1-Flash lie on the Pareto frontier of performance and cost among the evaluated models. Our code is available at https://github.com/lamda-bbo/agentic-bbo.
Hugging Face Daily Papers · Original source
Why is Anthropic running a biology lab?
Why is Anthropic running a biology lab? Anthropic is best known for Claude, the artificial intelligence chatbot that competes with OpenAI’s ChatGPT and Google’s Gemini. But the AI company now employs molecular biologists, studies DNA, and operates a laboratory where scientists conduct physical experiments. The San Francisco Bay Area facility is what researchers call a “wet lab”: a laboratory where scientists work directly with biological materials, rather than limiting their research to computer models and simulations. For an AI company, that may sound like an unusual expansion. Anthropic says there is a straightforward reason for it: if Claude is going to help scientists make discoveries about biology, those discoveries eventually have to be tested in the physical world. The company formed its life sciences research group in the spring of 2026 to investigate whether general-purpose AI models could systematically speed up biological discovery. Anthropic says building its own laboratory allowed the same team to work across the entire process, from improving Claude’s biological reasoning to testing its hypotheses experimentally. The project has already produced an early test case. Anthropic said on September 23 that Claude had identified a previously uncharacterised enzyme system associated with unusual repeating DNA sequences. The company says the pattern has some similarities to CRISPR systems, although researchers do not yet know what the newly identified system does. The finding has been released as a preprint and still requires further scientific investigation. Why does an AI company need a physical lab? AI can search scientific literature, analyse enormous datasets, write computer code, and propose explanations for patterns it finds. But biology ultimately depends on what happens inside cells, proteins, DNA, and other physical systems. A model might predict that a particular protein performs a certain function, for example, but scientists still need laboratory experiments to establish whether that prediction is correct. That creates a bottleneck for companies trying to use AI for science: computers can generate hypotheses far more rapidly than researchers can experimentally verify them. Anthropic’s lab is intended partly to close that gap. Its scientists use Claude to search large collections of genetic information for proteins or biological systems tha…
Discovery · Anthropic · Original source

AI4Science Daily · 2026-10-07 — Microsoft Research Podcast: What AI gets wrong and what failure teaches us

Wed, 07 Oct 2026 22:11:40 +0000

Selected developments

Microsoft Research Podcast: What AI gets wrong and what failure teaches us
Jennifer Neville is a partner research manager at Microsoft who’s built a career around understanding and advancing AI for real-world use, and much like the human-AI interactions she’s been studying, her early-career path was multiturn: math, then physics; cognitive science, then work; and finally computer science—despite her best efforts to avoid the field. In this conversation with Principal Applied Scientist Chad Atalla , she explores the role evaluation plays in pushing the performance boundaries of today’s AI systems to meet user needs and the “surprising failures” that emerge when models are tested beyond traditional benchmarks. Neville also shares practical guidance for working with current AI systems and discusses why looking closely at data matters when results defy expectations, and what decades of AI progress have taught her about predicting what comes next. From an unexpected career trajectory to the frontier of AI interaction and learning, this episode asks a larger question: what can we learn when the path—whether human or artificial—doesn’t unfold the way we expect? Subscribe to the Microsoft Research Podcast : Apple Podcasts Email Android Spotify RSS Feed Transcript [MUSIC] JENNIFER NEVILLE: There were times that I felt like I was spinning my wheels. Things weren’t working. I would get, kind of, dejected. And then we’d get to a point where we did learn something, and it was just the, the emotional thrill of that was, that was really what hooked me. The being able to, kind of, understand something that no one else understood yet … CHAD ATALLA: Sure … NEVILLE: … because it’s just at the frontier of what we know about things. STANDARD INTRODUCTION: This is the Microsoft Research Podcast , where Microsoft researchers—driving advancement through fundamental science and technology research—explore the who, how, and what’s next in computing and AI. [MUSIC ENDS] CHAD ATALLA: Hello, and welcome. I’m Chad Atalla, an applied scientist here at Microsoft Research. Today, I’m joined by Jennifer Neville, a partner research manager at Microsoft Research and the Samuel D. Conte chair professor of computer science and statistics at Purdue. Her research examines machine learning and AI for interactive domains and structured data, looking at how the data points that AI systems are trained on affect their behaviors and how that aligns with what users actually wa…
Microsoft Research AI for Science · Microsoft Research Feed · Original source
Hexagon
[This is a guest post by Ben Antieau , cross-posted from his blog . — T.] There was no personal problem, no world problem, whose eloquent solution did not exist— somewhere in some hexagon. —Jorge Luis Borges, “The Library of Babel” [1] New repository I am writing to announce Hexagon , a new repository for research works in the mathematical sciences and theoretical computer science. This has been a joint effort of many people over the last couple of months, starting in early August 2026. We felt that the proliferation of results proved with LLM assistance required a new model for collecting the associated papers. Besides wanting these works to be citable and reliably stored for long periods of time, we wanted to allow lower barriers for authorship and for community interest than the arXiv. We also wanted for these works to have a home besides random GitHub repositories or X/Bluesky/Mathstodon/LinkedIn posts. In order to support the development of this project, we founded the Hexagon Mathematics Foundation in the state of Delaware and we intend to apply for recognition of tax-exempt status under section 501(c)(3). The Board of Directors consists of Mohammed Abouzaid, François Charles, Bryna Kra, David Savitt, and Lauren Williams. I serve as the first Executive Director. And, we have a wonderful Advisory Board, where Kevin Buzzard, Akhil Mathew, Johannes Schmitt, Steinn Sigurðsson, Nikhil Srivastava, Ravi Vakil, and Rachel Ward serve. For more details, see our who we are page. My personal interest When working with a group of people brought together by an FRG grant, we clicked a button and discovered a new result, namely derived invariance of Hodge numbers of fourfolds in characteristic zero. In the past, this would be a result I would have submitted to an excellent, but not top, journal. In this case, there is no way I would put my name to the result. Instead, I blogged about it here . This is a brittle way to aggregate mathematics. Now, it is on Hexagon as hexagon:2609.00104 . The creation of new results happens naturally in working with LLMs. They bubble up, even as side results. My personal hope is that it becomes the community standard to identify these and submit them to Hexagon. The Lifecycle of Hexagon Objects (The title alludes to Ted Chiang’s “The Lifecycle of Software Objects” [2].) My ambition is that Hexagon is where mathematical ideas emerge from…
Terence Tao · Original source
OpenAI Releases 372 More Math Results as Skeptical Mathematicians Ask for Proof
The floodgates are open. Right on the heels of OpenAI’s new large language model producing the biggest math breakthrough in two decades, the company just released no less than 372 results from the same model. Each resolves or makes substantial progress on a major open question in mathematics or theoretical computer science, the company says. OpenAI revealed the results in a GitHub repository at 6 P.M. EDT. The deluge will take mathematicians months to parse through and understand—including to determine whether the proofs contain novel and important ideas or are mostly mash-ups of existing techniques. But many of the results have already been verified in Lean—a programming language that validates a proof’s logic—and are therefore all but certain to be correct. Among the results claimed are a solution to the four-dimensional Kakeya conjecture, improvements on some of the world’s most important computer algorithms, and actual progress toward math’s scariest problem, the Riemann hypothesis. On supporting science journalism If you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today. A spokesperson for OpenAI told Scientific American that the new model—which the company has not released to the public—produced almost every one of the results in response to a single prompt handed to a single AI agent. This would be a striking difference from OpenAI’s earlier blockbuster solution to the Navier-Stokes problem, which was produced through the collective efforts of a 10,000-strong agentic swarm that cost millions of dollars in computing power. If true, it would mean that unprecedented mathematical power could soon be accessible to anyone. But OpenAI’s reputation, among mathematicians, for bold claims and little transparency is causing considerable skepticism. “Until and unless they release the model and people can replicate their results, I think you should treat any claims about one-shotting problems with a single agent as unverified,” says Andrew Sutherland, a mathematician at the Massachusetts Institute of Technology. “We should ask for receipts.” (The OpenAI spokesperson also told Scientific American that some results might have taken multiple attempts.) OpenAI’s handling of the Navier-S…
Discovery · OpenAI · Original source
OpenAI releases 722 AI-generated mathematics papers as verification questions arise
OpenAI releases 722 AI-generated mathematics papers as researchers weigh what comes next OpenAI has published hundreds of mathematical manuscripts produced by an unreleased AI model, opening a much bigger debate over how AI-generated research should be checked, understood and incorporated into mathematics OpenAI has released a collection of 722 mathematical manuscripts generated by an internal frontier AI model, covering hundreds of research problems across mathematics and theoretical computer science. The collection is organized into 372 groups of related results and includes proofs, alternative arguments, consequences and supporting material. Many of the results also include formal proofs written in Lean, a programming language that allows mathematical arguments to be checked computationally. OpenAI says the work emerged from testing its internal models on open research problems after performance on its existing mathematics evaluations began to plateau. Around 4,000 problems were presented to the model during the evaluation process. OpenAI says the average result used computing resources equivalent to roughly three hours of ChatGPT Pro thinking. The scale of the release makes it considerably different from a typical AI benchmark. Rather than measuring whether a model can answer a fixed set of questions, OpenAI is publishing work that is intended to enter the mathematical research process itself. But the company is also explicit that publication does not mean every result has already been fully verified. The repository contains work at different stages of checking, and OpenAI says some manuscripts that have not yet been formalized in Lean could contain errors. Corrections and revisions will be recorded while earlier versions remain available. From solving problems to publishing research OpenAI first outlined its approach to releasing AI-generated mathematical results in September, when it said it was working with mathematicians on how such work should be presented responsibly. The company said it wanted to improve transparency around AI-generated mathematics by publishing supporting information alongside the results, including formal proofs where available, reasoning summaries and details about the computational resources used. Ten abridged summaries of the model’s reasoning have also been published for results spanning areas including number theory, compl…
Discovery · OpenAI · Original source
Machine learning for clinical time-series forecasting: an interview with Mayra Elwes
In a series of interviews, we’re meeting some of the PhD students that were selected to take part in the Doctoral Consortium at the International Joint Conference on Artificial Intelligence and the 29th European Conference on Artificial Intelligence (IJACI-ECAI 2026) . This time, we hear from Mayra Elwes about her work developing machine learning methods for clinical time-series data. Tell us a bit about your PhD – where are you studying, and what is the topic of your research? I am a second-year computer science PhD student at the Institute for Biomedical Informatics at the University Hospital Cologne. My research focuses on the generalisation and domain adaptation of machine learning for clinical time-series forecasting, especially for clinic-to-clinic generalization scenarios. I am working on practical strategies to evaluate methods in a more contextualised way and on improving existing domain adaptation methods specific to time-series forecasting. Could you give us an overview of the research you’ve carried out during your PhD? One major challenge in medical AI is performance degradation when models are deployed in clinics not represented in their training data. This performance degradation often stems from differences in patient demographics, clinical practice, or measuring devices, and it manifests as data distribution shifts. I address this problem for a use-case from the intensive care unit: forecasting blood oxygenation levels and arterial mean blood pressure based on observed vital parameters to improve hypotension and hypoxemia prediction. Since this application task is not covered by standard AI benchmarks, my clinical collaborators and I designed a new benchmark. Currently, we are collecting data at University Hospital Cologne. Tools and datasets I built to benchmark generalization for this vital parameter forecasting task are available as open-source software： WavePrep , Algorithm2Domain , MIMIC-III Waveform Database Matched Subset . While building this benchmark, I developed a strategy to benchmark generalisation and domain adaptation approaches for application tasks and domains in a more contextualised way. By reporting the shift relevant to the downstream task alongside the performance degradation, results become comparable even on private datasets and between different application tasks. Is there an aspect of your research that has been pa…
AIhub · Research news and interviews · Original source

AI4Science Daily · 2026-10-06 — Google Earth AI’s Population Dynamics Foundation Model applied to global public health

Thu, 08 Oct 2026 10:14:51 +0000

Selected developments

Google Earth AI’s Population Dynamics Foundation Model applied to global public health
October 6, 2026 Arbaaz Muslim, Software Engineer, Google Research, and Gautam Prasad, Software Engineer, Google Research With Google Earth AI’s Population Dynamics Foundation Model (PDFM), we can address the data gaps and temporal reporting lags of existing epidemiological workflows. In our latest work, we present five partner-driven case studies demonstrating how this model exemplifies the planetary geospatial foundation model paradigm for global public health. Public health decisions rely heavily on timely, granular data. For health conditions, such as cardiovascular disease or postpartum depression, that data shows us where to focus resources and support. For acute disease outbreaks, such as dengue or cholera, it can help inform urgent operational protocols and resource allocation. However, conventional epidemiological surveillance is often hindered by limitations in data: multi-year reporting lags, data siloed by rigid geopolitical boundaries, and data sparsity. Even without these limitations, traditional modeling approaches require extensive task-specific data collection and custom data engineering pipelines that are difficult to deploy during rapid outbreaks or in resource-constrained settings. To address these systemic bottlenecks, we introduce a new paradigm in public health leveraging planetary geospatial foundation models. Using Google Earth AI’s Population Dynamics Foundation Model (PDFM) as a proof-of-concept, we demonstrate how self-supervised, pre-trained representations of "place" can be integrated directly into existing health sciences and epidemiological workflows as plug-and-play inputs — enhancing the statistical and machine learning (ML) models epidemiologists already use, rather than building new pipelines from scratch. PDFM compresses privacy-preserving search trends, human mobility, built-environment density, and environmental determinants into location embeddings. Without requiring task-specific fine-tuning, these off-the-shelf location embeddings matched or improved on conventional inputs across a wide variety of disease domains, geographic settings, and epidemiological tasks. Part of Google Earth AI — our suite of geospatial models connecting satellite imagery, weather, anonymous search trends, human mobility, and other population dynamics — PDFM uses self-supervised learning to synthesize the following diverse, privacy-preserving…
Google Research · Earth AI · Original source
Erdős problems site founder Thomas Bloom on AI's gains and concerns
[This is a guest post by Thomas Bloom , crossposted from the Erdős problems forum . This blog post was initially written in a different file format and converted using AI. — T.] AI is changing everything, for better and/or worse, and the rate of change is dizzying; in recent months this has been particularly evident in mathematics, where AI has gone from being essentially useless to helping solve some of the hardest problems in mathematics in less than a year. The website http://www.erdosproblems.com has often been on the front line of these changes, and a barometer by which one could measure AI capabilities. This was not at all my intention when I created the site — I wanted to promote these problems to a human audience, and make a useful reference on what work has been done. But the easy availability of a large pool of questions that are simple to state, ranging from easy and obscure to deep and impenetrable, provided the ideal showcase for AI. In some ways this has led to a huge amount of progress — we now know the answer to many questions we did not before. (Although a lot of this recent progress has come from increased human efforts and renewed interest in some problems, rather than just AI solutions.) There have also been negative effects, however: some mathematicians have dismissed Erdős-style mathematics as ‘easy/recreational’; many have stopped thinking about Erdős problems believing that they cannot compete with AI; and, most significantly, there has been a wave of AI-produced solutions provided with no explanation. While of course these are useful and tell us new things, they are also displacing and discouraging those who are actually interested in the mathematics. (These issues have stopped being Erdős-specific as AI continues to make breakthroughs in an ever wider range of fields.) A couple of weeks ago I asked for feedback on how people used the site, and what they thought should be changed. I received, both publicly and privately, a wide variety of opinions, and I’m grateful to everyone for their feedback. What was particularly striking was the number of people I heard from who have benefited a lot from the site, learning new mathematics and finding new problems to think about, but have never commented on the site. I have been particularly conscious of this large silent, audience, and have tried to make sure their experiences were not drowned…
Terence Tao · Original source
OpenAI's Navier–Stokes Claim Puts Mathematicians in a Reckoning at Heidelberg Forum
On 13 September, leading luminaries and up-and-coming talents in mathematics and computer science congregated at the annual Heidelberg Laureate Forum in Germany for a week of discussion, networking, and—naturally—wurst. Having attended several of these events in the past, the mathematicians normally focus their chatter on which famed researchers are present, or what interesting problems they have been working on. But instead, every snippet of conversation caught in passing or any debate accidentally overheard was about how AI companies such as OpenAI, Anthropic, and Google are steamrollering their way through mathematics. And there is good reason for these fervent discussions. Mathematics is the perfect testing ground for AI, involving step-by-step logical reasoning and answers that are automatically and objectively verifiable. This has led tech giants to develop their AI mathematical capabilities at a terrifying rate this year, leading to both new solutions to previously unsolved problems and a reckoning within the community as to what it means to do math in the age of AI. Tech Giants Target the Millennium Problems In short order, AI has gone from struggling with everyday research-level problems to solving a whole raft of teasers posed by prolific Hungarian mathematician Paul Erdős, then verifying the proof of Fermat’s Last Theorem, and—most recently and famously—OpenAI announcing it had solved the Navier-Stokes existence and smoothness problem. As attendee and young researcher Ailsa Robertson, of the University of Amsterdam, put it: “AI and LLMs set the math community on fire over summer.” Of the seven extremely difficult Millennium Prize Problems posed in the year 2000 by the Clay Mathematics Institute, only one, the Poincaré conjecture, has been solved by humans so far. OpenAI’s claim to have solved a second, the Navier-Stokes existence and smoothness problem, if verified, represents a watershed moment for automated reasoning. Fields Medalist Jacob Tsimerman, of the University of Toronto, was keen to recognize this achievement during a press conference at the forum: “We’ve seen rapid increases in capabilities, even faster than many people, including myself, expected,” he said. “Is AI doing something new? I don’t know the details, but…solving Navier-Stokes feels pretty definitive.” Since this announcement, rumors have swirled about which of the remaining…
Discovery · OpenAI · Original source
Sherpa: Teaching LLMs to Teach Adaptively
Large language models (LLMs) have become increasingly capable problem solvers, but being able to solve a problem is not the same as being able to teach it. Existing approaches to training LLMs as teachers rely on demonstrations, preference data, or predefined pedagogical criteria that specify what good teaching looks like. However, these signals are often not grounded in individual student learning outcomes, where effective teaching strategies can vary substantially across learners. To address this, we introduce Sherpa, a multi-turn reinforcement learning framework that instantiates multiple student archetypes with LLMs conditioned on distinct learning preferences and trains a teacher model to adapt its instruction by directly maximizing their learning outcomes. Teacher LLMs trained with Sherpa improve instructed students' performance across all archetypes by an average of 20.5 percentage points. Under MathTutorBench's evaluation, Sherpa raises the overall pedagogy score from 52.5% to 79.2%, indicating better teaching responses. Our human studies show that the trained teacher is preferred over the base model in 79.6% of pairwise comparisons. Together, Sherpa trains LLM teachers to adapt to diverse simulated students and become better aligned with human teachers, paving the road towards AI tutors teaching real students.
Hugging Face Daily Papers · Original source
insitro's Machine Learning Models for Small Molecule in vivo Pharmacokinetic Prediction Now Available in Lilly TuneLab
Share SOUTH SAN FRANCISCO, Calif. – October 6, 2026 – insitro, the AI company unlocking causal human biology to systematically generate high-impact therapeutics, today announced that the advanced machine learning models that insitro built with Eli Lilly and Company (Lilly) to predict in vitro and, critically, in vivo properties of small molecules can now be used by certain biotech companies in Lilly TuneLab™. Trained on multi-species preclinical data that Lilly generated over decades of research from hundreds of thousands of unique molecules, the models offer a powerful alternative to today’s industry standard method of extrapolation from in vitro assays, providing an enterprise-grade capability traditionally available only to organizations with the scale to generate such data volumes themselves. The models will be updated as the dataset continues to expand. Announced in September 2025, the collaboration paired insitro’s computational and AI/ML expertise with Lilly’s proprietary preclinical dataset of in vitro and in vivo measurements from a wide array of compounds with established ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) properties, derived from decades of Lilly drug discovery, and representing a world-class dataset in quality, consistency and scale. This collaboration demonstrates the power of insitro’s platform by combining insitro’s AI and ML expertise with Lilly’s data to unlock and advance small molecule development. “Nine in 10 programs that enter the clinic fail, most often when we first test for efficacy. That happens when the mechanism did not drive the disease, but also when the molecule never reaches the tissue at the right concentration for the right duration,” said Daphne Koller, Ph.D., founder and CEO of insitro. “The question of a molecule’s in vivo behavior has been an empirical question, answered experimentally, slowly, one compound at a time. These models make it a computational question first.” The downstream implications extend to patients. Most small molecule candidates fail, often for pharmacokinetic or safety reasons that surface late, after significant time and capital have been committed. Better prediction at the design stage may mean that fewer candidates advance on flawed structures, promising molecules reach the clinic on shorter timelines, and program development becomes more tractable. Improved in…
insitro · News · Original source

AI4Science Daily · 2026-10-05 — Interview with William Yijiang Li: vision language models and the physical world

Tue, 06 Oct 2026 22:13:57 +0000

Selected developments

Interview with William Yijiang Li: vision language models and the physical world
At the International Conference on Machine Learning (ICML 2026) , Dezhi Luo, Yijiang Li, Maijunxian Wang, Tianwei Zhao, Bingyang Wang, Siheng Wang, Pinyuan Feng, Pooyan Rahmanzadehgervi, Ziqiao Ma , and Hokin Deng presented their work Vision Language Models Cannot Reason About Physical Transformation . In this interview, William Yijiang Li tells us more about the research, the method the team used, and the controlled experiments they carried out. What is the topic of the research in your paper and why is it an interesting area for study? Our research asks a fairly fundamental question: do vision-language models actually understand how the physical world changes over time? Modern vision-language models are increasingly being used in areas such as robotics, embodied agents, and video understanding. They can recognize objects, describe scenes, and answer surprisingly sophisticated questions. However, being able to describe a scene is different from maintaining an understanding of an object or physical quantity as the scene changes. We study this through the concept of conservation . For example, if water is poured from a short, wide glass into a tall, narrow one, its appearance changes dramatically, but the amount of water remains the same. Similarly, spreading a row of coins farther apart does not change the number of coins. These tasks are simple for humans, but they require an important form of reasoning: you need to track what happens during a transformation and distinguish changes in appearance from changes in the underlying physical state. This is particularly interesting for embodied AI. If an AI system is expected to interact with the real world, it cannot rely only on static visual recognition. It needs to maintain and update representations of objects and their properties as events unfold. What are the main contributions of your research? I would summarize the contributions in three parts. First, and most importantly, our study shows that current vision-language models still struggle to maintain and update a stable representation of the physical world as it changes over time. Although these models can often answer basic questions about physical concepts such as conservation correctly, our controlled experiments show that this does not necessarily mean they understand the underlying transformation. Models that perform well when a quantity is conserved…
AIhub · Research news and interviews · Original source
Google DeepMind creates watermarked AI-designed proteins that still work in lab tests
Google DeepMind creates watermarked AI-designed proteins that still work in lab tests SynthID Bio embeds a detectable signature into AI-generated protein sequences and structures, with Nature research showing the watermark can survive without significantly disrupting biological function Google DeepMind researchers have demonstrated a way to watermark AI-designed proteins while preserving their biological function, an attempt to bring the provenance systems already used for AI-generated media into synthetic biology. The new method, SynthID Bio, puts a detectable signature directly into either a protein’s amino acid sequence or its predicted three-dimensional structure. The central challenge is very different from watermarking an AI-generated image. Changing a few pixels may leave a picture effectively unchanged. Altering a protein can change how it folds, binds to other molecules or functions altogether. DeepMind’s researchers had to show not simply that their watermark could be detected, but that the resulting biological designs still worked. In laboratory testing reported in Nature, watermarked protein binders remained functional across three targets: VEGF-A, PD-L1 and the receptor-binding domain of the SARS-CoV-2 spike protein. Pushmeet Kohli, Chief Scientist at Google Cloud and VP Science at Google DeepMind, described the result on LinkedIn as the “successful synthesis of AI-designed proteins that are both functional and watermarked.” How do you watermark a protein? SynthID Bio actually covers two different approaches. For protein sequences, the system subtly changes how amino acids are selected while an AI model generates the sequence. The resulting signature can then be detected using a secret watermarking key. For three-dimensional protein structures, the researchers took a different route. They fine-tuned part of AlphaFold 3 so the watermark is built into the predicted atomic coordinates produced by the model. The recommended version of that structural method achieved a detection rate above 99.8% at a 0.1% false-positive rate in the researchers’ tests, without reducing key structural accuracy measures compared with the AlphaFold 3 baseline. The more important test for the sequence watermark happened outside the computer. Researchers synthesized watermarked and non-watermarked protein binders and tested their ability to attach to the three target prot…
Discovery · Google DeepMind · Original source
Detectrons convert transient RNA sequences into stable DNA barcodes for high-throughput analysis of RNA-dependent processes
Programmable RNA sensors such as toehold switches are used to detect specific RNA sequences. However, their reliance on protein-based or RNA-based outputs limits their use in multiplexed and sequencing-based applications. Here we introduce Detectrons, modular biosensors that couple programmable toehold switches with retron-mediated reverse transcription to transduce RNA inputs into unique DNA barcodes, converting dynamic RNA signals into durable DNA records within living cells. The framework enables alternative modes of transcript-based sensing with applications including viral infection detection. Through the construction of a synthetic toehold retron library and application of machine learning, we uncover key design principles that improve signal strength and specificity. We apply Detectrons to the multiplexed live-cell detection of specific phage infections, enabling transcript-triggered barcode synthesis and quantitative host susceptibility profiling in pooled bacterial populations. Detectrons provide a scalable and generalizable strategy for phage screening and for recording transcriptional events in complex bacterial communities. RNA sequences are detected and converted directly into DNA barcodes.
Nature Portfolio · Nature Biotechnology · Original source
Photonic neuromorphic learning via generalized in situ physical gradient descent
Physics-based neuromorphic computing promises major gains in performance and energy efficiency, but training remains largely in silico, requiring costly physical modeling and suffering from fabrication-induced errors. Here we introduce on-chip, in situ physical gradient descent (INSPIRE), a general training method for photonic integrated circuits. INSPIRE uses on-chip synthetic time-reversal holography to measure the full complex fields of bidirectional photonic modes, enabling gradient computation and parameter updates directly in the physical system. The framework is topology-agnostic and compatible with diverse optical circuits. Experiments yield trained matrices with a relative error of 0.26%, and we demonstrate in situ training through scattering media for matrices larger than the number of native tunable elements. INSPIRE also enables meta-photonic circuits to perform in situ meta-learning, realizing single-shot photonic learning with 251-fold model compression and 136-fold task-specific training acceleration. This work offers a practical route toward adaptive and efficient intelligent photonic systems. This study reports a physics-based in situ learning method that trains artificial neural networks directly within the photonic integrated circuits, turning diverse optical circuits into efficient AI learning hardware.
Nature Portfolio · Nature Computational Science · Original source
FairRSFM: A Biome-Aware Benchmark and Debiasing Framework for Remote Sensing Foundation Models
Remote sensing foundation models (RSFMs) are commonly evaluated using aggregate metrics, which can hide systematic performance disparities across ecological regions. We introduce FairRSFM, a biome-aware benchmark for evaluating ecological group robustness in RSFMs. FairRSFM maps georeferenced samples from 14 terrestrial biome classes into six ecologically meaningful macro-groups and evaluates models under a unified frozen-backbone evaluation protocol. The benchmark covers four downstream datasets: m-EuroSAT, m-BigEarthNet, m-SA-Crop-Type, and MMEarth20K with Dynamic World label maps. Using Prithvi-EO-2.0, SatMAE, and DOFA across three random seeds, we show that aggregate performance consistently masks biome-dependent disparities across architectures and tasks. For example, Prithvi-EO-2.0 reaches 90.98% overall macro-F1 on m-EuroSAT but a mean worst-group score of only 83.72%, while m-SA-Crop-Type drops from 27.30% overall mIoU to 18.47% in the Xeric and Mineralogical group. We further evaluate Biome-Orthogonal Linear Probing (BOLP), Dynamic Biome Reweighting (DBR), and GroupDRO as complementary mitigation baselines. Their effectiveness is model- and task-dependent; for example, BOLP improves Prithvi-EO-2.0 worst-group F1@opt on m-BigEarthNet from 46.12% to 50.27% without updating the RSFM backbone. FairRSFM provides a reusable protocol for diagnosing and mitigating ecological robustness gaps in remote sensing foundation models. Code and datasets are available at: https://github.com/aminurhossain/FairRSFM.
Hugging Face Daily Papers · Original source

AI4Science Daily · 2026-10-04 — Classical Solutions and Singularity Formation in Incompressible Fluids

Wed, 07 Oct 2026 10:16:11 +0000

Selected developments

Classical Solutions and Singularity Formation in Incompressible Fluids
[This is a guest post by Diego Córdoba and Luis Martínez-Zoroa . This blog post was initially written in a different file format and converted using AI. — T.] Euler formulated the equations for an inviscid incompressible fluid more than 250 years ago. Published in his 1757 memoir , they belong to the earliest systems of partial differential equations in mathematical physics, following d’Alembert’s work on the wave equation. Understanding the evolution they describe has required the work of many generations of mathematicians. One of the central questions is whether an initially regular fluid flow can develop a singularity in finite time. The recent work on Euler and Navier–Stokes obtained with the aid of large language models has brought renewed attention to this question. These exciting developments build on decades of mathematical research, including works from recent years, when the field has continued to be specially active. In this post we describe some contributions to that body of knowledge, including our own. We will go over some of that recent mathematical literature, and discuss the ideas of constructing blow-ups through a cascade of vortex layers. Partial differential equations provide a common language for problems ranging from wave propagation and fluid motion to geometry and general relativity. For an evolution equation, the Cauchy problem asks us to determine the future from prescribed initial data. A local existence theorem gives a solution for a short time; the next question is whether it continues for all time or develops a singularity at a finite time. This distinction between global existence and finite-time blow-up is already visible in the elementary Riccati equation The solution starts from a finite value and remains smooth until , when it becomes unbounded. The quadratic nonlinearity in three-dimensional Euler suggests a similar possibility, but the fluid problem also involves transport, geometry and a pressure determined nonlocally by the entire flow. The challenge is to understand whether the full equation can sustain the amplification suggested by this simple ODE. The central aim of our research has been to construct classical solutions of incompressible fluid equations that develop singularities in finite time, starting within a regime of local existence and uniqueness. We want the initial data to determine a unique regular evolut…
Terence Tao · Original source
ASCENT: Online Test-Time Training of Long-Horizon Agents via Self-Distillation of Verified Experience
A large language model (LLM) agent solves long-horizon tasks through many reasoning-action turns, with one verification signal at termination. Deployed agents face streams of related tasks, making their trajectories a natural resource for improvement. In-context adaptation agents store reflections, memories, or skills as text, so reuse depends on retrieving the right experience and on a frozen policy executing it. We study Online Agentic Test-Time Training (OaTTT), which trains the LLM's weights on its own execution trajectories during deployment. The agent executes each task once, in one pass over the stream, and the executed trajectory with its verification result is the only learning signal for weight updates that persist across tasks. Directly imitating or reinforcing the generated tokens of this single attempt destabilizes the policy. We introduce ASCENT (Agentic Self-distillation for Cross-task EvolutioN at Test-time), which instead self-distills verified experience. A stable version of the LLM, its frozen initial copy, receives the verified trajectory as privileged information and predicts next-token distributions along it with this hindsight. Distilling them into persistent LoRA fast weights updates the agent for later tasks, without an external reference solution or stronger teacher. By further removing invalid-action turns, ASCENT distills enhanced privileged experience for more efficient execution. We characterize its population target and the limits of sparse outcome selection. Across ALFWorld, WebShop, and AppWorld at varied model scales, ASCENT improves task success and interaction efficiency as experience accumulates, outperforms online adaptation methods, and transfers to held-out scenes, showing that an agent can consolidate verified experience into its weights without a separate training phase or memory retrieval. Project page: https://artificer-ai-lab.github.io/ASCENT
Hugging Face Daily Papers · Original source
LLM-as-Jev: LLMs Are Already Jev-Style Decision Models -- When and How to Fine-Tune Them
Jev-style decision models return categorical probability distributions over predefined options without generating free-form text, enabling software systems to act on their outputs directly. In this work, we investigate the extent to which general-purpose LLMs already possess this capability out of the box, and when fine-tuning is actually necessary. We present LLM-as-Jev, an architecture-preserving framework that extracts calibrated decisions directly from next-token probabilities over bracketed numeric identifiers. LLM-as-Jev provides both a training-free inference recipe and a fine-tuning objective that optimizes candidate selection via a tree-factorized listwise loss while anchoring auxiliary predictions to the base model using KL divergence penalties. Evaluating on Qwen3.5-4B and Qwen3-0.6B, we find that modern LLMs are inherently effective decision models: without training, the 4B model matches community Jev-style models built on the same backbone, outperforms letter-logit readouts, supports arbitrary option counts, and natively handles multimodal decisions over images. Fine-tuning provides targeted rather than universal benefits -- substantially improving weaker models and specific tasks (such as many-option intent routing), but offering diminishing returns for strong backbones. Crucially, our KL anchors prevent behavioral degradation in conversational text generation, with LoRA delivering the strongest performance on capable models.
Hugging Face Daily Papers · Original source

AI4Science Daily · 2026-10-03 — Solving Open Research Problems Together: Six Papers from Meta AI and Mathematicians

Thu, 08 Oct 2026 03:01:09 +0000

Selected developments

Solving Open Research Problems Together: Six Papers from Meta AI and Mathematicians
Solving Open Research Problems Together Mathematicians and Muse Spark collaborate on six research papers Earlier this year, our models achieved gold-medal-level performance across five high-school Olympiad competitions in mathematics, physics, and chemistry. Those results inspired us to see whether AI could also help scientists solve open research problems. Competition problems can be incredibly difficult, but those problems already have a solution. Open research is different. There is no answer key, no guarantee that an approach will work. Making progress means trying and retrying ideas, making new mistakes and resolving them, and sometimes going back to the beginning. Over the past several months, we've partnered with mathematicians to explore problems across several areas of mathematics. They used both Muse Spark 1.1 and 1.2 in Thinking Mode through the regular meta.ai chat interface, with no custom research scaffold. Open research takes time. Our goal here wasn't to mass-produce papers, but to empower researchers and help them develop mathematical insights that others can understand and build on. We’re proud to partner with the scientific community to help them advance research. These collaborations were done under the following principles: - A team of mathematicians guided the research and worked with Muse Spark to explore ideas and develop the arguments. - A second group of mathematicians then reviewed their work. - Each paper clearly marks which passages were primarily drafted by researchers and which were drafted by AI. - Each paper gives credit to the earlier research and mathematical ideas it builds on. Today, we're sharing six such papers from that collaboration. Five present answers to previously open research questions. After completing our work, we learned that other teams outside Meta had independently announced solutions to some of the same problems using different approaches. We recognize and appreciate their contributions and clearly acknowledge their work and how it relates to ours in the papers. Probability: The Strict Threshold for Gaussian Ellipsoid Fitting Aykut Arslan with Muse Spark via meta.ai. Review by Babak Modami, Alexander Roitershtein, Mark Sepanski, and Grigory Sokolov Paper: The Strict Threshold for Gaussian Ellipsoid Fitting. The paper answers a long-standing question about fitting random Gaussian points in high dimensions…
Discovery · Meta AI · Original source
LoHi: A Joint Allocation Perspective on Frames, Pixels, and Front-End Latency in Long-Video Understanding
Efficient long-video understanding with vision-language models (VLMs) is often framed as selecting informative frames or visual tokens at a fixed native resolution. We show that per-frame resolution can instead be traded for denser temporal coverage, while front-end decoding latency depends on the size of the candidate pool rather than the final token budget. An empirical study across multiple VLMs and long-video benchmarks yields three findings: dense low-resolution sampling outperforms sparse native-resolution sampling at matched token budgets; resolution-sensitive tasks benefit from selected high-resolution frames; and front-end decoding dominates wall time for hour-long videos. Motivated by these findings, we introduce LoHi, a training-free, single-pass framework that combines a dense low-resolution video stream with sparse high-resolution image frames through the VLM's native video and image pathways. LoHi-Anchor selects high-resolution frames using codec I-frame metadata, while LoHi-SemDiv uses query relevance and visual diversity over CLIP features. Across three long-video benchmarks, LoHi improves average accuracy by 10.6 percentage points over the native-resolution baseline at a matched token budget and by 5.2 percentage points over the strongest prior efficiency method. It also reduces front-end decoding latency by up to 7x on hour-long videos. Project page: https://sixundong.com/projects/lohi
Hugging Face Daily Papers · Original source
LMBuild: Evaluating LLM Agents for Generating Buildable and Functional Structures
LLM-based agents are increasingly capable of generating complex 3D structures, with the potential to reshape how objects are designed and realized in the physical world. Yet, producing elegant geometry is fundamentally different from producing objects that can be built and perform their intended functions. Existing evaluations largely focus on geometric quality while overlooking physical realizability. We introduce LMBuild, a benchmark for evaluating LLM agents on generating buildable and functional structures. LMBuild represents generated objects as assembled structures comprising part decompositions, joints, materials, and sequences. To support reproducible evaluation, we provide a unified framework consisting of: (1) an interactive environment in which agents can use tools to retrieve, create, and place components to construct objects; (2) a curated benchmark that repurposes established CAD datasets and augments them with knowledge from Wikipedia; and (3) a evaluation framework covering structural soundness, functional affordance, design quality, and physical realization. Evaluations across 30 systems reveal several intriguing findings: (a) Soundness and alignment are no longer the primary bottlenecks for frontier closed-source models, while functional affordance and physical operability remain substantially more challenging; (b) stronger models more effectively create new components, whereas weaker models tend to rely on retrieval; and (c) providing functional specifications substantially improves part completeness, kinematics, and physical operability. These results show that generating real-world structures requires deeper reasoning about functional affordances, mechanics, and designing and creating novel components. We expect LMBuild to provide a foundation for measuring progress and incentivizing research toward agents that generate buildable and functional structures.
Hugging Face Daily Papers · Original source
The Numerical Linear Algebra of Large Language Models
Numerical Linear Algebra (NLA) has consistently played a vital role in advancing science by providing tools to solve fundamental problems encountered in scientific and engineering applications. Over the decades, it has continually evolved to meet the demands driven by successive waves of scientific discovery. For instance, during the 1950s and 1960s, substantial efforts were devoted to developing methods for solving eigenvalue problems that emerged from the rapidly growing field of aerodynamics. This led to the discovery of the LR and QR algorithms. Later the attention turned to the solution of sparse linear systems that were common in applications like computational aerodynamics. Today we are experiencing yet another wave of major scientific advancement and NLA is once more at the heart of its development. This Machine Learning (ML) wave is proving to be utterly disruptive in science and engineering. Many tools in ML particularly Large Language Models (LLMs) are grounded in matrix and tensor methods. As we are approaching Artificial General Intelligence (AGI), it is clear that matrix methods will be called to play an even more significant role. For the numerical linear practitioner the speed of the current change makes it particularly challenging to adapt. This is a survey article that centers on machine learning techniques, with a particular focus on large language models. It has two main objectives. The first is to clarify the core concepts behind Large Language Models in a manner accessible to specialists in numerical methods. The second is to examine the key Numerical Linear Algebra concepts employed by LLM techniques, while also highlighting several significant recent contributions of NLA to the field.
Hugging Face Daily Papers · Original source

AI4Science Daily · 2026-10-02 — Google DeepMind watermarks AI-designed proteins, and the first binders still bind

Mon, 05 Oct 2026 10:15:15 +0000

Selected developments

Google DeepMind watermarks AI-designed proteins, and the first binders still bind
Google DeepMind watermarks AI-designed proteins, and the first binders still bind Google DeepMind has taken its SynthID watermark into biology, and says the first watermarked protein binders still bind. In wet-lab testing against three target proteins, VEGF-A, the SARS-CoV-2 spike protein RBD and PD-L1, the watermarked designs “matched the hit rate, binding affinity, and natural sequence diversity of unwatermarked versions”, the company says. It calls those “the first-ever watermarked and biologically functional protein binders”. The method is published as a peer-reviewed paper in Nature, “Function-preserving watermarking of AI-generated proteins”, and Google describes the work as a proof of concept. The mark goes into the sequence, not into a file SynthID Bio “embeds an imperceptible signature directly into the biological code”, Google says, so the watermark is “verifiable not just on a digital model but on the synthesized, physical protein itself”. A design that leaves the model and gets made in a laboratory carries the mark with it. For binders, molecules built to latch onto other proteins, the team used AlphaProteo, DeepMind’s own binder design method, alongside a SynthID Bio-enabled version of ProteinMPNN, a widely used protein sequence generator. Adaptyv Bio helped with the in vitro validation. AlphaFold 3 carries the watermark in its weights The second target is 3D structures. Google says SynthID Bio fine-tunes “a small part of AlphaFold 3’s diffusion network” so that predicted coordinates “inherently carry a detectable signature regardless of who runs the model”, and says prediction accuracy is preserved alongside near-perfect detectability. Google gives two reasons for wanting any of this. Novel AI designs “can bypass traditional DNA synthesis screening”, and mislabelled synthetic 3D structures “risk polluting public databases and misleading downstream research”. Tampering is the part that does not work yet “Moving forward, key challenges include making the watermark more robust against deliberate tampering,” Google says. It suggests pairing the watermark with provenance metadata, along the lines of C2PA for digital media, or with central repositories of AI-generated biological data. The work on viruses is further back. With the Hie lab at Stanford University and the Arc Institute, DeepMind put SynthID Bio into the genomic model Evo 2 and watermark…
Discovery · Google DeepMind · Original source
Open-sourcing AstaBrief 8B, a small open model for fast cited report generation
Language models can already help researchers search the literature, synthesize evidence, and work through complex questions. But scientific work places particular demands on these models—answers need to stay grounded in evidence, the models need to preserve what the evidence actually supports rather than quietly broadening a study’s conclusions, and researchers need to be able to verify the final outputs. We see that in how scientists use Asta, our agentic platform for scientific work. Instead of simple keyword searches, users often bring substantial context and many constraints—for example, asking Asta to compare approaches across a body of literature while accounting for a particular method, population, or setting. Many also return to generated reports later, treating them as working research artifacts rather than one-off answers. We wanted to help scientists generate cited reports faster, with a model they could download and run themselves. To do that, we tested whether a small, open model trained specifically for scientific report generation could match the report quality of the proprietary models we were using, while reducing generation time and serving costs. We built AstaBrief 8B, a model that turns a research question and retrieved literature excerpts into a cited report. AstaBrief is available in Asta’s Generate a report feature today as Fast mode alongside Claude-powered Thinking mode, and we’re also open-sourcing it and the training data so others can study, reproduce, and build on our approach. Developing AstaBrief required tens of thousands of real research queries, citation-focused filtering, preference data, and a redesigned report-generation pipeline that writes the full report in one pass rather than section by section. The result is nearly an order-of-magnitude reduction in report generation time compared to the proprietary models we tracked—across the full Asta pipeline, Fast mode averages 51.1 seconds per report compared with 178.5 seconds for Thinking mode, about 3.5× faster. Together, those efficiency gains made AstaBrief a useful test case for a broader goal: building open language models that can be adapted to the specific demands of scientific work. Open weights will also let institutions run AstaBrief on their own infrastructure, which is necessary when research questions reveal sensitive or unpublished work. Alongside the model we…
Allen Institute for AI · AI for Science · Original source
Google DeepMind develops invisible watermarks for AI-designed proteins
October 1, 2026 report Google DeepMind develops invisible watermarks for AI-designed proteins Paul Arnold Author Gaby Clark Scientific Editor Robert Egan Senior Editor Watermarks have long protected everything from bank notes and fine art to digital photographs and software, helping prove authenticity and trace an object's origin. Their main function is to protect copyright and verify authenticity. That same traceability is needed in synthetic biology, where AI is now a useful research tool. To address this, Google DeepMind has introduced SynthIDBio, a method for embedding invisible signatures directly into biological sequences and 3D structures without interfering with their function. AI is actively designing new functional proteins and predicting their three-dimensional shapes. But along with the benefits come a host of potential problems. These include biosecurity risks, such as the potential misuse of AI-designed biological molecules, and the spread of fake or misleading scientific data. The team released details of its protein watermarking in a paper published in the journal Nature. How to watermark a protein There are two kinds of watermarks, digital patterns built directly into the protein design data. Which one is used depends on what the AI is generating. If the model is writing a protein sequence, the watermark is a secret code hidden in the order of the amino acids. If the model is predicting a 3D biomolecular structure, the watermark is an invisible pattern hidden in the 3D coordinates of the atoms. Then came the jump from computer models to the real world. Researchers made watermarked protein binders in the lab to target three specific molecules. These included part of the coronavirus spike protein, a human protein involved in blood vessel growth and a human protein involved in regulating immune responses. They then needed to find out two things. The first was whether the watermarks changed how the proteins work, and the second was whether a detector could spot them. The watermarked proteins attached to all three targets with comparable strength to nonwatermarked versions. Meanwhile, the detection tool correctly identified all of the watermarked protein sequences and more than 99.8% of the watermarked 3D structures. Adding the watermarks did not significantly affect how well the proteins worked or the accuracy of the 3D structures. There are st…
Discovery · Google DeepMind · Original source
Pivot-SD: Efficient Self-Distillation for Masked Diffusion Language Models
Masked diffusion language models (dLMs) offer a promising parallel alternative to autoregressive models for complex reasoning. However, they face a distinct credit-assignment challenge, since a few commitments during denoising sharply reduce the uncertainty over the remaining masked positions and shape much of the response. Most post-training recipes for dLMs do not use this signal to decide which tokens to train on: they typically train on the final text or assign rewards to whole denoising steps, rather than selecting the individual commitments that shape the response. We introduce Pivot-SD, an efficient offline self-distillation framework that supervises only these high-impact commitments (pivots). Pivot-SD selects pivots using an information-gain metric measuring uncertainty reduction over the remaining masked positions. Pivots from successful trajectories are trained with cross-entropy, and pivots from failed trajectories with targeted unlikelihood, leaving the rest of the failed trajectory untouched. Using only 200 questions and four rollouts each, Pivot-SD improves LLaDA-8B-Instruct over full-sequence SFT and budget-matched diffusion RL baselines across math and code benchmarks.
Hugging Face Daily Papers · Original source
SEA-Driven OdorNet: A Large-Scale Standardized Molecular Olfactory Dataset Based on a Hybrid Semantic Taxonomy
The development of machine olfaction (digitization of odor) is currently hindered by data scarcity and the lack of standardization in descriptive labels. Datasets integration faces the challenges of descriptor-level conflicts, including cross-source labeling inconsistencies, synonym redundancy, skewed distributions, etc. To address the semantic confusion inherent in multi-source data, we developed a SEA (Statistical co-occurrence, Expert correction, and AI alignment) taxonomy framework that integrates statistical co-occurrence analysis, expert knowledge (perfumery taxonomy), and AI-assisted semantic alignment based on the definitions from both encyclopedia and odor-related scientific literatures. A unified-labeled olfactory dataset, namely OdorNet, is further presented, which is by far the largest molecule dataset with unified olfactory labels comprising over 20,000 entries and 9,000 unique molecules originated from academic literatures (1960–2021) and open databases. This process consolidates cluttered labels into a hierarchical system with the most extensively effective vocabulary coverage in the field to date. Technical validation demonstrates that a baseline model on OdorNet achieves a Macro F1 score of 0.42, substantially higher than the scores obtained on commonly used odor datasets and on OdorNet itself using the same model but without the SEA-based label taxonomy. This indicates that OdorNet allows to better capture learnable structure-odor relationships, laying a solid fundamental for digital odor editing and human-computer interaction.
Nature Portfolio · Scientific Data · Original source

AI4Science Daily · 2026-10-01 — Google DeepMind publishes SynthID Bio, a watermarking method for AI-designed proteins

Mon, 05 Oct 2026 10:15:15 +0000

Selected developments

Google DeepMind publishes SynthID Bio, a watermarking method for AI-designed proteins
In what can be called a major breakthrough for synthetic biology and AI-powered biosecurity, Google DeepMind, Google’s AI lab, has developed a new method to watermark AI-designed proteins, establishing a permanent proof of origin directly inside biological code. Published in the journal Nature, the proof-of-concept technology, which is named SynthID Bio, solves a long-standing technical challenge: embedding an non-noticeable, verifiable digital identifier into biological sequences and predicted molecular structures without degrading their physical real-world utility in laboratory tests. The rollout comes amid rapid advancements in biological machine learning, where systems like AlphaFold 3, AlphaProteo and ProteinMPNN allow scientists to predict complex folding patterns and engineer functional binders. Yet these generative capabilities have heightened global biosecurity risks. Unfamiliar, machine-generated sequences can slip past conventional gene-synthesis filters, while unverified synthetic structural data risks contaminating open-access scientific repositories. ‘An urgent challenge for the AI era’ Google DeepMind Chief AI Scientist Demis Hassabis emphasised that guarding against biological misuse must be treated as a frontline safety priority for frontier AI: “Biosecurity is one of the most urgent challenges for the AI era. Bringing SynthID to biology so AI-generated proteins can be watermarked is a critical step - and we’re open sourcing SynthID Bio tools so the research community can build on this work. Published in @Nature today, congrats to the team!” Pushmeet Kohli, Vice President of Science at Google DeepMind and Chief Scientist at Google Cloud, celebrated the practical synthesis of these molecules: “Very happy to announce that our team at Google DeepMind has pushed the boundaries of generative biology, achieving the successful synthesis of AI-designed proteins that are both functional and watermarked. This proof-of-concept watermarking of the building blocks of life is enabled by SynthID Bio, our new protein watermarking method. It is designed to safeguard the new era of AI-powered generative biology and strengthen global biosecurity.” How SynthID Bio operates SynthID Bio functions as a modular toolkit tailored across two primary dimensions of structural biology: SynthIDBio-sequence that subtly steers the selection of amino acid building blocks du…
Discovery · Google DeepMind · Original source
Machine Learning Reveals That Elephant Choruses Encode Behavior
Image 1. A matriarch leads her family across the dry lake bed in Amboseli, Kenya. Such a movement is typically preceded by single and chorused Let’s go and Cadenced rumbles as individuals initiate and negotiate a plan of action. Image Credit: ElephantVoices . Key Takeaways Around 60% of elephant rumbles occur in overlapping choruses , yet these multi-caller exchanges have been far less studied than isolated calls. In a new study published in Scientific Reports and authored in collaboration with ElephantVoices , we analyzed elephant choruses as they naturally occur, without needing to separate each call . We could predict broad behavioral context from the choruses much better than chance. We also identified vocal convergence during some coordinated interactions, with some rumbles becoming progressively more acoustically similar as an exchange unfolded. Elephants face habitat loss and human-wildlife conflict across their range. Understanding how elephants communicate, make decisions, and coordinate their movements could help protect key habitats and corridors , ease human-elephant conflict , and support people and elephants living alongside one another. The Challenge of Studying Elephant Choruses In many social contexts, including when elephants greet each other after time apart, several animals may rumble at once. Some of those low-frequency, overlapping sounds even extend into the infrasonic range below human hearing. Roughly 60% of elephant rumbles occur inside choruses of two or more voices. Yet these exchanges have received far less scientific attention than isolated calls because they are difficult to analyze. That means a large part of elephants’ natural communication has not been studied. Instead of trying to pull each individual voice apart, we wanted to know whether we could predict behavior from choruses as a whole. In a new study published in Scientific Reports , working with ElephantVoices and Amboseli Trust for Elephants , we used machine learning to extract behavioral information directly from complex, overlapping elephant choruses. Listening to the Whole Chorus We analyzed 903 rumbles from adult females recorded between 1986 and 2022. The recordings come from one of the most closely studied elephant populations in the world. Since the 1970s, researchers from the Amboseli Elephant Research Project have followed these families across generations…
Earth Species Project · Blog · Original source
A multimodal synthetic 3D point cloud dataset of shoe uppers for robotic processing boundary segmentation
Abstract Robotic processing of shoe uppers requires accurate segmentation of thin processing boundaries in 3D point clouds. However, acquiring and annotating large collections of real-world scans is costly. We present a multimodal synthetic 3D point cloud dataset of shoe uppers for robotic processing boundary segmentation. The dataset contains 1,067 samples, including 1,000 synthetic point clouds and 67 real-world test scans for independent evaluation. The synthetic data generation pipeline includes texture and appearance augmentation, region-aware geometric deformation, and sensor degradation simulation. Both branches provide three-dimensional coordinates, RGB attributes, and pointwise binary processing boundary labels. The synthetic branch additionally includes panel identifiers, while the real-world labels were manually refined. Technical validation with standard point cloud segmentation backbones assesses dataset usability under a controlled synthetic-to-real evaluation protocol. The released point clouds form a complete, directly usable dataset. The accompanying code and assets support regeneration and adaptation of synthetic subsets.
Nature Portfolio · Scientific Data · Original source
Introducing SynthID Bio: Google DeepMind Puts Watermarks Inside AI-Designed Proteins
Introducing SynthID Bio: Google DeepMind Puts Watermarks Inside AI-Designed Proteins Google DeepMind is introducing SynthID Bio after laboratory tests produced the first watermarked protein binders that retained their intended biological function. The September 30 announcement moves AI watermarking beyond images, text, audio, and video. Its signature can remain detectable in a synthesized physical protein, not merely in a digital design file. That makes the work more consequential than another content-labeling experiment. The central conflict is provenance versus control. SynthID Bio can identify outputs from participating models, but it cannot establish that every unmarked protein is natural or safe. Researchers also found that determined users can remove its sequence watermark by redesigning the protein. The result is a credible proof of concept with a demanding path to deployment. It pressures model developers, biological databases, and DNA synthesis providers to decide whether shared provenance standards belong in the infrastructure of AI-assisted biology. Introducing SynthID Bio as a Physical Provenance Test Google DeepMind has shown that an AI-generated protein can carry a detectable signature without losing the function researchers designed it to perform. SynthID Bio is a family of methods for watermarking protein sequences and predicted biomolecular structures. A watermark is a hidden statistical signal that authorized software can detect later. It indicates an output came from a compatible AI system. The project has two distinct components. SynthIDBio-sequence influences which amino acids a sequence-generation model selects. SynthIDBio-structure introduces small, detectable patterns into the atomic coordinates predicted by a folding model. That distinction matters because a protein sequence and a predicted structure represent different things. The sequence specifies an ordered chain of amino acids. The structure describes how atoms in that chain are arranged in three dimensions. For the sequence test, researchers combined AlphaProteo-designed backbones with a modified version of ProteinMPNN. ProteinMPNN generates amino acid sequences expected to fold into a supplied protein structure. The team evaluated binders against three targets: VEGF-A, PD-L1, and the receptor-binding domain of the SARS-CoV-2 spike protein. Protein binders are molecules design…
Discovery · Google DeepMind · Original source
ScholarCatalyst: A Benchmark for Retrieving Papers That Inspire New Research
What makes great scientists great? Even as AI systems start to make progress on open problems, scientists remain far ahead of them at sensing which prior idea, buried in an ever-growing archive of research, a new problem needs. To study this skill, we draw on researchers who know firsthand which earlier work advanced their completed projects, with papers serving as pointers to the ideas within. Using our automated pipeline that makes author annotation scalable, we build ScholarCatalyst by having 184 lead authors of 207 recent computer science papers label which candidates did or could have advanced their project, each with a detailed rationale. We introduce a retrieval task with author-provided judgments: given an initial research question, retrieve these papers from only the literature available when the project began. Agentic search does no better than embedding retrieval (0.42 vs. 0.48 Recall@20) despite calling that same retriever as a tool. Even an agent built on Claude Fable 5.1, which may have seen the completed papers during training, reaches only 0.51 R@20. These results highlight the need for new training recipes that equip models with expert intuition for searching broad corpora. We envision ScholarCatalyst as a step toward scientific agents that can take a half-formed idea and point to the prior research it needs.
Hugging Face Daily Papers · Original source

AI4Science Daily · 2026-09-30 — Microsoft Research Debuts Quine, a Multimodal World Model of Biology

Mon, 05 Oct 2026 22:15:24 +0000

Selected developments

Microsoft Research Debuts Quine, a Multimodal World Model of Biology
AI Models & Platforms Microsoft Research Debuts Quine, a Multimodal World Model of Biology Add Unite.AI to your preferred sources on Google Microsoft Research on September 29, 2026, introduced Quine, a research effort built around a multimodal world model of biology and an interactive harness that links the model to orchestration and reasoning models, scientific tools, the research literature, the wet lab, and the teams of scientists using them. With the announcement, Microsoft opened applications for the first Quine Fellows cohort and reported that, working with researchers at the Broad Institute of MIT and Harvard, it has used the system to rank compounds predicted to produce therapeutic shifts in tumor cell state, with several top-ranked candidates validated across multiple wet-lab assays. The announcement is authored by Nicolo Fusi, VP and distinguished scientist, and Jonathan M. Carlson, vice president. A World Model Built Across Biological Scales Microsoft defines the world model it is pursuing as one that represents the state of a biological system, forecasts how that state would change under an intervention, and reasons through the downstream consequences several steps ahead. The company says such a system would not replace experimentation; instead, it would let researchers weigh and rank candidate directions computationally before committing limited laboratory resources. Quine’s core model is trained to learn representations shared across biological scales and data types, among them sequence, structure, function, cellular state, and imaging. Its training modalities span genomics, proteins, chemistry, RNA and cell state, and bioimaging. Microsoft reports that joint training across these modalities strengthens rather than dilutes performance: evidence from one modality can inform predictions in another, capturing relationships the company says could stay siloed or be lost if separate single-domain specialist models were orchestrated instead. In Microsoft’s description, the system runs as a cycle that starts and ends with the scientist: a question yields proposals, proposals turn into experiments worth running, and the resulting measurements refine both the next question and the model itself. The model need not be perfect, the company says, adding that it will never perfectly capture biology; it needs only to guide experimental design usefully. The Qu…
Discovery · Microsoft Research AI for Science · Original source
CytoVI: deep generative modeling of antibody-based single cell data
Antibody-based single-cell technologies, such as flow cytometry, mass cytometry and CITE-seq, have become widely used in clinical diagnostics and basic research; however, their analysis is complicated by technical noise, batch effects, platform differences and restricted antibody panels. Here we present CytoVI, a probabilistic generative model for statistically rigorous unified analysis of antibody-based single-cell data. CytoVI generates informative cell embeddings, imputes missing measurements, performs differential protein expression testing and automates annotation of cells in a single probabilistic model. We applied CytoVI to build an integrated B cell maturation atlas spanning 350 proteins and identified proteins associated with immunoglobulin class-switching. In a cohort of patients with B cell non-Hodgkin lymphoma profiled by flow cytometry and CITE-seq, CytoVI uncovered disease-associated T cell states. CytoVI is available as open-source software at scvi-tools.org .
Nature Portfolio · Nature Methods · Original source
Does Learning Protein Folding Generalize to Broader Reasoning?
Large language models rely heavily on human text, which often conveys surface answers rather than the spatial and structural logic behind them. Protein folding is a natural testbed, because one solved structure yields thousands of exactly checkable spatial and topological statements. We ask: can learning to fold proteins teach general models reusable reasoning capabilities? To answer this, we build FoldingCorpus, a protein-derived question-answer dataset, and Fold2Reason, a recipe that post-trains on it through two complementary signals: discrete structural answers predicted via the model's native language head, and continuous 3D geometry decoded from the same shared representations. On FoldBench, Fold2Reason achieves structure prediction scores 2.7 to 3.5 times those of Qwen3.5-9B. Beyond protein structure prediction, it improves performance on all 10 benchmarks spanning spatial, graph, scientific, and general reasoning, raising macro-average accuracy from 45.09% to 48.33% (+3.23 pp), with positive gains on all 10 benchmarks, while matched controls built from random, synthetic, and shuffled structure yield substantially smaller or negative gains. Our work shows that non-linguistic, structure-dense scientific data can systematically improve broad reasoning in language models, making a solved scientific problem a practical source of post-training supervision.
Hugging Face Daily Papers · Original source
JevSpawn: Adaptive Agentic Inference through Compositional Action Spaces
LLM agents generate intermediate reasoning and actions token by token, making extended interactions slow and computationally expensive. Jev-style models offer fast probabilistic predictions over finite fields, but require those fields to be specified in advance. This requirement limits autonomous task solving, where the available actions must be derived from natural language instructions and adapted through interaction. We introduce JevSpawn, a compositional policy that connects natural language task specifications to finite probabilistic exploration. Parallel action spawning is coupled with feedback driven branch selection, representation revision, and recovery from retained alternatives. Shared action structure and model prefixes reduce repeated generation and context computation without additional training. Evaluations on eight benchmark tasks against seven agent baselines and a TypeSafe Jev variant establish JevSpawn as a promising approach to structured agentic inference, with improved task performance and faster navigation.
Hugging Face Daily Papers · Original source
OpenAI's Noam Brown on multi-agent systems: 10,000 agents deserve at most 10% of the credit for the Millennium Prize breakthrough
Noam Brown, an OpenAI researcher and core author of o1, discussed his current focus, multi-agent systems, in a recent podcast. He said OpenAI previously used 10,000 agents over 88 hours and 130 billion tokens to solve a Millennium Prize mathematics problem. Brown was explicit that the result came from having a general, powerful model, and that he would not even attribute 10% of the credit to multi-agent methods, which he feels have received disproportionate praise because they are eye-catching. Multi-agent systems simply scale test-time compute in parallel rather than serially, at some efficiency loss since agents do not share full context. His testing reached only 16 agents: four agents solve a problem about twice as fast for roughly double the cost, and 16 still improve performance with slightly lower efficiency. Speedups are somewhat sublinear and domain-dependent, and ablation experiments at 1,000-agent scale would be unaffordable.
量子位 · Original source

AI4Science Daily · 2026-09-29 — Joint Statement about Mathathon

Thu, 08 Oct 2026 03:01:09 +0000

Selected developments

Joint Statement about Mathathon
[This is a guest post by a coalition of Caltech mathematicians: the original organizers of Mathathon (Alvan Arulandu, Andrea Li, Avni Garg, Brian Zhao, Caiman Moreno-Earle, Sathvik Redrouthu), two coauthors of the Open Letter about the Mathathon (Dylan King, Tasmin Chu), and other members of the Caltech community (Mayla Ward, Merrick Hua, Robert Joseph George, Sergei Gukov, Shaowu Zhang, Tony Yue Yu, Vihaan Dheer). This blog post was initially written in a different file format and converted using AI. It is also crossposted at Proofs and Prompts — T.] Mathathon is centered around one question: how can we use AI to augment human understanding of mathematics? This past week, the original organizers of Mathathon, two coauthors of an open letter about the mathathon , and other members of the Caltech community engaged in a conversation to redesign this event. We quickly reached a consensus: problem solving is but one component of mathematics; mathematical understanding and exposition are similarly meaningful. But these components have often been overlooked by grantmakers and hiring committees. We want to celebrate human mathematicians who undertake this work. Thus the new theme of Mathathon is Old Problems, New Proofs . Consider the four color theorem, the ABC conjecture, or the Navier-Stokes problem. Many mathematicians find their proofs—or claimed proofs—difficult to understand or unsatisfying. We invite our participants to pick a problem with an unintuitive solution, learn as much as possible in 40 hours, and present their findings to their peers. Then, they will take two months to develop an alternative proof or exposition. Afterward, they’ll submit an explainer and a GitHub repository. The explainer can take any form: a paper, a blog post, a video, an interactive game, et cetera. The repository will store whatever was used to produce the explainer—LLM chat histories, code used to produce visualizations, and more—so the mathematics community can examine the ideas behind the finished product. By shifting the focus from open problems, we want to encourage Mathathon participants to explore everything else we do to make sense of mathematics: generalizations, new notations and techniques, connections to other problems. These innovations can be more exciting than solving individual problems. Mathathon treats AI as one of many tools available to mathematicians. We…
Terence Tao · Original source
Introducing Quine: An AI research system designed for the complexity of biology
Biology doesn't operate in silos, and neither should the AI representation of it. Quine is an early-stage research effort to create a multimodal world model of biology. By connecting insights across biological scales and modalities, Quine helps scientists computationally search a space far larger than intuition allows and prioritize hypotheses before they reach the lab. Experimental results provide important feedback, helping researchers sharpen future research directions.
Microsoft Research AI for Science · Microsoft Research Feed · Original source
Building a new path to make medicines with AI - Isomorphic Labs
Listen: Solving all disease is one of the most monumentally complex challenges in science. Today, for the first time, AI provides the foundation to make it achievable. Tackling the frontier of human health is our mission, and it means identifying exactly where modern AI can make the greatest difference. We see drug discovery as the critical wedge in solving all disease, and therefore our initial focal point where AI can have a transformative impact on human health. Developing a drug that successfully and effectively treats disease fundamentally involves understanding biological mechanisms and designing exquisitely engineered molecules that can modulate them safely. So we are taking a straight shot at making that a reality. We are generating the data, training the AI models, developing the software, and laying down the physical infrastructure to make new medicines; taking these potential therapeutics all the way to patients in the clinic. We are building this new architecture in a way that will scale across multiple disease areas, and different types of medicines, from small molecules to antibodies, peptides, and beyond. In 2021, when we started Isomorphic Labs, AI drug discovery was just a hypothesis. Conventional approaches were bottlenecked by slow, resource-intensive iterative design-make-test cycles, viewing AI as an experimental tool for narrow tasks. At the same time the AI revolution was well underway across multiple industries. A seismic shift was happening with the emergence of foundation models, moving away from single-task algorithms toward highly versatile systems capable of world-understanding, zero-shot reasoning and generating complex text and images. Up to this point, I had spent my education and career in core AI, building intelligent systems by scaling up neural networks and deep reinforcement learning to solve the hardest problems at a superhuman level. But having grown up in a medical family, I recognised the potential to apply this superhuman new technology to humanity’s most meaningful challenge: human health. Working with Sir Demis Hassabis, we created the founding team at Isomorphic Labs to pivot breakthroughs in AI toward solving the drug discovery problem. We set out on a path to approach drug discovery as an engineering process, stripping back to first principles, to rebuild the process to design medicines. Almost every field of e…
Discovery · Isomorphic Labs · Original source
Hidden Dates in System Prompts Shift LLM Evaluation Results
Reproducibility is essential for scientific research, yet prior work shows that LLM outputs vary with hardware and batching. We identify an overlooked factor: the hidden injection of the current date into system prompts, which users cannot control and which changes every day. Across 9 recent LLMs and 6 datasets spanning multiple-choice QA (MCQA), math reasoning, code generation, and machine translation, performance varies solely with the current date, with deltas of up to 6% on MCQA, 14% on math reasoning, 7% on code generation, and 2.84 BLEU on machine translation. Model rankings also shift, affecting leaderboards. This date effect exceeds other sources of non-determinism, such as batch size and numerical precision. Standard prompting techniques -- chain-of-thought and few-shot prompting -- do not reduce the sensitivity; chain-of-thought even amplifies it. Our findings underscore the need for careful evaluation protocols to ensure reproducibility and fair comparisons in LLM research.
Hugging Face Daily Papers · Original source
RLE-Bench: A Qualifying Exam for Coding Agents as Robot Learning Engineers
Coding agents are beginning to move beyond purely digital tasks to tackle physical-world challenges, particularly in robotics. Existing robotics benchmarks, however, primarily focus on the performance of individual artifacts, such as policies or controllers, offering limited coverage of coding agents' broader engineering capabilities. Real-world robotics extends beyond control: agents must build, integrate, diagnose, and improve heterogeneous artifacts under resource constraints and reason from multimodal feedback. To evaluate these broader capabilities, we introduce RLE-Bench, a benchmark of robot-learning tasks spanning four representative robotics development workflows: interactive control, policy learning, perception and estimation, and mechanical design. We use diverse task-specific metrics to evaluate the artifacts submitted by the coding agents, from the success rate the agents achieved to the policy agents trained, the harness agent built, and the mechanical structures the agent designed. We aggregate these metrics into an overall RLE Index and report workflow-specific capability profiles, enabling systematic comparison of coding agents' capabilities across multiple capability dimensions. Beyond performance ranks, we also conduct in-depth case studies examining agent behavior on representative tasks, highlighting both current capabilities and limitations, and pointing to the opportunities robotics tasks have to offer for future agent training.
Hugging Face Daily Papers · Original source

AI4Science Daily · 2026-09-28 — Microrobot navigation policies trained in under 10 minutes

Sun, 11 Oct 2026 01:44:51 +0000

Selected developments

Microrobot navigation policies trained in under 10 minutes
Microrobots hold considerable potential for various applications in which targeted navigation is a basic requirement. Deep reinforcement learning has recently emerged as a powerful paradigm for fully autonomous microrobot navigation. Yet, current deep reinforcement learning-based approaches pay limited attention to learning efficiency and effectiveness, requiring hours to days for model training. Consequently, this impedes both rapid practical deployment and parameter optimization. To address these challenges, we present a learning framework that enables effective microrobot navigation policies to be trained within minutes. Here in the proposed framework, we develop a fully vectorized simulator with more than 10,000 artificial vascular environments, parallelizing dynamics, ray-casting-based visual feature extraction and feasibility checks across thousands of environments to achieve roughly 190,000 transitions per second. To achieve effectiveness in the fast training, we propose a task-shaping-regularization reward framework. The task-shaping-regularization framework accelerates convergence, improves final performance, reduces action variation by at least 33.7% and increases obstacle clearance by at least 2.1% across all evaluated scenarios. Results show that the proposed learning framework reduces training time to under 10 minutes, while supporting zero-shot deployment across distinct microrobot types and navigation scenarios. Collectively, this framework can substantially shorten the design loop and accelerate the deployment of autonomous microrobots. A vectorized simulator and structured reward framework enable microrobot navigation policies to be trained within minutes and transferred without retraining across robots and environments.
Nature Portfolio · Nature Machine Intelligence · Original source
AI-powered platforms predict proteins that undergo phase separation
From left: Dr Liang Qiyu, research fellow at the Institute for Digital Molecular Analytics and Science at NTU, and Professor Miao Yansong. Credit: NTU Singapore. Scientists at Nanyang Technological University, Singapore (NTU Singapore) have developed two artificial intelligence (AI)-powered platforms that could enable researchers to more accurately predict proteins that undergo phase separation – a process through which proteins in cells separate like oil droplets in water. The tools are the result of a systematic analysis of predicted phase-separating proteins from various living things, including animals, plants, fungi and single-celled organisms such as bacteria. From their analyses, the scientists also uncovered fundamental insights about the phase separation process. Led by Professor Miao Yansong from the School of Biological Sciences and the Institute for Digital Molecular Analytics and Science (IDMxS) at NTU, in collaboration with Professor Weibo Gao from NTU’s School of Electrical and Electronic Engineering, the research could be applied to advance the understanding of diseases and ageing, as well as to improve crops. Both AI platforms and their research findings have been reported in the peer-reviewed journal Molecular Plant . Predicting which proteins separate Phase-separating proteins separate from their surroundings and concentrate into droplet-like compartments, enabling cells to bring molecules together, accelerate or suppress reactions to respond rapidly to stress. When phase separation is altered, these proteins can form abnormal condensates or aggregates, which are associated with diseases, including cancer and neurodegenerative disorders such as Alzheimer’s disease. In plants, changes in phase separation are linked to impaired stress responses and lowered immunity against microbial pathogens. Identifying proteins that undergo phase separation and determining how this behaviour is regulated are important for applications ranging from biomedical research on diseases and ageing to devising approaches for sustainable agriculture. However, accurate prediction remains challenging because phase-separating proteins assemble dynamically into complex structures and interact with numerous cellular components. Addressing these challenges requires comprehensive experimental datasets and systems-level information across diverse organisms. In addition, m…
AIhub · Research news and interviews · Original source
CheatBench: A Benchmark for Measuring Reward Gaming in AI Agents
Reinforcement learning has helped AI agents solve increasingly difficult tasks, but high rewards do not always reflect the work users intended. In recent incidents and controlled evaluations across the AI industry, agents trained to maximize reward have accessed unauthorized information, attempted to evade monitoring systems, and even breached sandbox protections to attack external systems. As agents become more capable, this behavior could pose increasingly serious risks. To measure this problem, we introduce CheatBench, a benchmark of cheating in AI agents across mathematical research, knowledge work, coding, visual tasks, and other domains. Its environments combine challenging assignments with opportunities to cheat, allowing researchers to study how agents pursue a goal when honest work is difficult. CheatBench supports comparisons across models and task categories, providing a testbed for measuring and reducing cheating as agents take on more consequential responsibilities. We publicly release CheatBench at https://cheatbench.ai
Hugging Face Daily Papers · Original source
MM-ABC: Towards Generalist Mobile Manipulation via Seeing, Coordinating and Imagining
Mobile manipulation extends robot interaction beyond a fixed kinematic workspace by making the reachable region itself controllable. This flexibility introduces two central challenges: spatially grounded perception under continuous ego-motion and coordinated control of heterogeneous arm and base actions. Existing approaches strengthen geometry through explicit 3D representations or predictive world models, and often decouple mobility and manipulation into separate action streams. We argue that effective mobile manipulation requires not only decoupling, but also representations that support efficient cross-stream collaboration. We present MM-ABC, a foundation model built around Seeing, Coordinating, and Imagining Arm-Base Collaboration. MM-ABC combines sparse multi-level VLM features for spatial perception; a training-only future branch that uses world imagination and geometric intent as extra supervision, strengthening perception and manipulation-intent prediction and improving the overall learning signal; and MM-APT, which coordinates separate manipulation and mobility streams through masked joint attention and clean-action x-prediction. In controlled ablations, replacing clean-action prediction with velocity prediction lowers success on RoboCasa365 composite-seen tasks from 32.8% to 29.2%, and removing future supervision or multilevel conditioning causes larger drops. We pretrain MM-ABC on 5,000+ hours of heterogeneous robot data spanning 400K+ episodes, 12 datasets, and 17 embodiments. Experiments cover EBench, RoboCasa365, ManiSkill-HAB, LIBERO, LIBERO-Plus, and real-world mobile manipulation. MM-ABC achieves 44.71% success on EBench, 61.2% on RoboCasa365, 99.1% on LIBERO, 82.8% on LIBERO-Plus without perturbation training, and 83% mean success on five real-world tasks.
Hugging Face Daily Papers · Original source
Unifying AI-assisted scientific discovery around exploration, hypothesis generation, and testing
In this review, Hasib et al. introduce EXHYTE a process centric survey to unify research on AI-assisted scientific discovery. Mapping recent studies reveals rapid advances in evidence retrieval, knowledge assembly, and idea generation, but it also reveals a persistent bottleneck in translating hypotheses into executable tests and using results to refine subsequent steps. The survey underscores the need for human oversight, reproducibility, validation, and auditable end-to-end discovery workflows.
Cell Press · Patterns · Original source

AI4Science Daily · 2026-09-27 — Unmask the State: When Does State Adaptation Matter for Masked Diffusion Language Models

Thu, 08 Oct 2026 03:01:09 +0000

Selected developments

Unmask the State: When Does State Adaptation Matter for Masked Diffusion Language Models
Masked diffusion language models (MDMs) admit flexible generation orders, making the unmasking strategy an inference decision. Existing methods vary in how they prioritize positions, control parallelism, restrict selection regions, revise predictions, or plan future denoising, yet it remains unclear when these choices should change during generation. We study this question through strategy reversals, where an alternative action becomes preferable to a fixed choice. We organize MDM inference into five axes--score, cardinality, region, commitment, and planning--and define adaptation opportunity as the one-step utility advantage of the best candidate action over a validation-selected fixed action. This view shows that adaptation value depends on both the frequency and magnitude of such reversals. Across three MDMs and ten tasks, adaptation opportunities are highly heterogeneous, with some regimes exhibiting concentrated and predictable one-step gains. This motivates selective adaptation: lightweight detectors calibrated on validation prompts identify high-opportunity states, capturing, for example, 56.9 percent of the candidate-set oracle opportunity by adapting only the top 10 percent of states on LLaDA-8B constrained JSON filling. Our transition-level results suggest that state adaptation is most useful when applied selectively rather than uniformly.
Hugging Face Daily Papers · Original source

AI4Science Daily · 2026-09-26 — Anthropic Says Claude Computed a Nine-Loop Particle Physics Amplitude

Fri, 09 Oct 2026 10:11:40 +0000

Selected developments

Anthropic Says Claude Computed a Nine-Loop Particle Physics Amplitude
AI Models & Platforms Anthropic Says Claude Computed a Nine-Loop Particle Physics Amplitude Add Unite.AI to your preferred sources on Google Anthropic on September 25, 2026 published a guest post on its research site reporting that two of its physicists used the company’s Claude model to compute the six-particle scattering amplitude in planar N=4 super Yang-Mills at nine loops, answering a public challenge issued to AI companies on August 7, 2026. The Challenge Matt von Hippel, a former theoretical physicist who now writes about physics at 4gravitons.com, set the terms in an August 7, 2026 blog post. He asked AI companies to show that an AI, using the computing resources an academic has access to, could solve one of the scattering amplitudes field’s big outstanding problems: determining whether N=8 supergravity diverges at seven loops, or finding the six-particle amplitude in N=4 super Yang-Mills at nine loops. “Give us N=8 supergravity to seven loops, or N=4 super Yang-Mills to nine loops,” he wrote. Von Hippel argued that these problems are hard in a computational sense: each loop represents an increase in complexity, in calculations that typically scale exponentially or even factorially in the number of loops. In principle, he wrote, amplitudes researchers could solve any of them with no new ideas, using known methods, but they would need access to far more computing power. At the end of August 2026, Liam Fitzpatrick and Siddharth Mishra-Sharma, two physicists at Anthropic, contacted von Hippel to say they had tackled one of the challenges, and after verifying the result with longtime amplitudes researcher Lance Dixon they walked him through the work. The post’s disclosure states that Anthropic invited von Hippel to write the guest post and compensated him for his time, that Anthropic staff gave feedback on drafts, and that the content and opinions are his own. The Physics Problem Scattering amplitudes are formulas that let physicists use the momenta and energies of subatomic particles to calculate how likely they are to react in particular ways. They are hard enough to compute that physicists almost always work with approximations cut off at a specific number of loops: the more loops included, the closer the result comes to the real answer and the harder the calculation becomes. Most scattering amplitude formulas have been calculated only to two loops,…
Discovery · Anthropic · Original source
Yes, Claude can do Nine Loops
Science In this guest post, physicist and science writer Matt von Hippel shares what happened when he issued a challenge to AI companies regarding a problem in his former subfield of theoretical physics. It’s not often that you issue a challenge, only to see it beaten a month later. But we’re living in unusual times. Let me introduce myself: I’m Matt von Hippel. I used to be a theoretical physicist; these days I’m a science writer. Throughout, I’ve been a blogger, writing weekly at 4gravitons.com about physics and the people who do it. More and more, blogging about physics has meant blogging about AI. That’s a problem, because I’m definitely not an AI expert. I’ve dabbled in it, sure. I probably know more than your grandma. But I mostly have to step back and trust the experts. And frustratingly, the experts disagree! I’ve heard from smart, well-informed people who are confident that AI is a few years away from superintelligence, and that superintelligence will be capable of truly terrifying things. And I’ve heard from smart, well-informed people who are equally confident that LLM-based AI is close to a ceiling, that models like Claude won’t even be able to do impressive work in physics, let alone conquer the world. I’ve been reluctant to make my own predictions. Before forming an opinion, I wanted to see an LLM make progress on something familiar, something I knew was hard to do because I’d tried to do something similar myself. In addition to that, I wanted to see an LLM do something that I expected to be computationally hard. LLMs have made impressive strides in math, certainly, and this month alone has likely changed many people's minds. But progress in math comes from new ideas, and ideas are mysterious things: one never quite knows how hard they are to find until they’re found. Computation felt more solid. I wanted to see an LLM tackle a challenge that seemed out of reach not because researchers didn’t know how to do it in principle, but because doing it seemed like the kind of thing that would take more computers and time than the researchers reasonably had access to. I wanted to see if those researchers were wrong: if a smarter, artificial researcher could use the same computers, and solve the problem anyway. So, I issued a cha…
Anthropic · Science Blog · Original source
OpenTumorBoard: A Real-World Benchmark of Multidisciplinary Tumor Board Discussion Trajectories
Multidisciplinary tumor boards integrate multimodal clinical observations and longitudinal patient histories through specialist discussions, yet benchmarks rarely capture these real-world trajectories. We introduce OpenTumorBoard, a benchmark with 611 patient cases and 19,157 discussion turns across ten specialist roles, transcribed from 12,534 minutes of publicly available tumor board recordings on YouTube. The benchmark evaluates two settings: SPECIALIST TURN, in which an LLM responds to a clinically significant question posed during a real discussion, and BOARD SIMULATION, in which it generates an entire back-and-forth discussion and reaches a consensus on therapy recommendations, surgical plans, next actions and clinical trial matching. Evaluation of 14 general-purpose frontier and medical LLMs reveals substantial limitations: the best models score 3.43 out of 5 in clinical equivalence to specialist answers and 2.78 out of 5 in alignment with recorded board conclusions. Supervised finetuning and reinforcement learning improve performance on a held-out test set, suggesting that real-world discussion trajectories can support model adaptation. Three M.D. experts review a subset of the benchmark, finding high information coverage and factuality of patient cases and strong fidelity of extracted consensus conclusions. We will release OpenTumorBoard and its automated curation pipeline to support the development and evaluation of LLMs for multidisciplinary, personalized cancer decision-making.
Hugging Face Daily Papers · Original source
LANTERN: Illuminating Hidden Mathematical Knowledge in Language Models
Language models can now prove theorems, but people still decide which problems to pursue. We ask whether a model's internal representations can help identify promising mathematical connections. We develop LANTERN, a fast, cost-efficient pipeline that uses a classifier over pretrained-model activations to rank candidate relations, followed by staged filtering, hypothesis generation, executable verification, and analytical checking. Applied to the On-Line Encyclopedia of Integer Sequences (OEIS), LANTERN ranked 50 million pairs among 10,000 frequently referenced sequences and produced 62 verified relations between pairs without an existing OEIS cross-reference. A content screen retained 13 relations worth presenting; nine of these are informative or insightful, including four which are entirely novel to the best of our knowledge: none appears in the OEIS or in our targeted literature search. The entire end-to-end process including classifier training, candidate ranking, filtering and verification took under 8 hours.
Hugging Face Daily Papers · Original source

AI4Science Daily · 2026-09-25 — Lila Sciences' AI Lab Identifies Six Palladium Catalyst Families for Green Hydrogen

Mon, 12 Oct 2026 11:40:19 +0000

Selected developments

Lila Sciences' AI Lab Identifies Six Palladium Catalyst Families for Green Hydrogen
AI Models & Platforms Lila Sciences’ AI Lab Uncovers Palladium Catalysts for Green Hydrogen Add Unite.AI to your preferred sources on Google Lila Sciences said on September 25, 2026, that its AI-directed laboratory proposed, synthesized and screened 2,942 catalysts for green hydrogen in three months, identifying six palladium-based material families for the acidic oxygen evolution reaction at the heart of hydrogen production. The research team detailed the work in a preprint submitted to arXiv on September 24, 2026: “AI-guided high-throughput discovery of iridium- and ruthenium-free palladium-oxide catalysts for durable acidic oxygen evolution,” by Ken J. Jenewein and 20 other authors. Producing hydrogen from water requires two half-reactions: the hydrogen evolution reaction at the cathode and the oxygen evolution reaction, or OER, at the anode. OER runs at high potentials in strongly acidic environments that dissolve most metals. In today’s commercial electrolyzers, iridium oxide is the catalyst best able to survive those conditions while remaining active; ruthenium is a less stable option. Both are among the rarest elements on Earth, and nearly all iridium is produced as a byproduct of platinum mining, at only a few tonnes per year worldwide. The Palladium Result While screening 2,942 oxides across 53 material systems and 26 elements, the platform began synthesizing and testing palladium compositions. According to the company, palladium had been considered a dead end for OER because it was assumed to underperform in acidic conditions. The model’s candidates were modified palladium compositions rather than the pure metal, with small additions of other elements transforming the metal’s behavior. During the second and subsequent campaigns, the program identified six palladium-based material families on or near the Pareto front for OER, meaning an optimal balance of activity and stability. After repeated activity tests and more than 1,000 hours of stability testing, Lila said the top metal composition performed comparably to ruthenium, an industry-standard OER catalyst. John Gregoire, the company’s chief autonomous science officer, said his former group at Caltech explored catalysts for this reaction for 12 years and that he would have advised a student proposing this combination of elements to pursue a more promising direction. “This combination defies all t…
Discovery · LILA Sciences · Original source
Science Podcast: How AI language evolves inside swarms, how baby orangutans learn to forage, and an astrophysicist on loss
First up on the podcast, Staff Reporter Tom Howarth joins host Sarah Crespi to talk about the language artificial intelligence agents evolve inside their swarms . Next on the show, producer Elah Feder talks with Revathe Thillaikumar , a behavioral ecologist and postdoctoral fellow at the Max Planck Institute of Animal Behavior. The two discuss the value of social learning as baby orangutans learn to find their own food over the longest dependency period of any animal, except humans. Finally, from our series on scientific autobiographies, books host Angela Saini talks with astrophysicist Sara Seager about the moving account of life after losing her husband in her book, The Smallest Lights in the Universe . This week’s episode was produced with help from Podigy . About the Science Podcast Image: inkoly/istock Learn more about your ad choices. Visit megaphone.fm/adchoices
Science · Science Podcast · Original source
When compression techniques don’t just add up: Interaction effects in hybrid LLM compression
What happens when you shrink a language model twice — once by removing some of its weights, once by storing what’s left with fewer bits? You’d guess the damage just adds up. It doesn’t. The problem: does compression just add up? Small language models are increasingly used for on-device apps that need to run in tight memory budgets [Zhou et al. 2025]. To shrink them enough, people often combine two tricks: pruning, which removes some of the model’s weights, and quantization, which stores the remaining weights using fewer bits. Recent work has combined both, often finding the mix beats either trick alone [Kang et al. 2025]. But most of these studies only look at the combined result — they don’t ask whether the two tricks are simply adding their effects together, or doing something else entirely. We wanted to find out. A simple test: does it just add up? To check, we borrowed a simple idea from statistics. For each benchmark, we measure four things: the original model, the pruned-only model, the quantized-only model, and the hybrid. Then we compare: If the two tricks were simply adding up, would land close to zero. If is positive, the hybrid does better than expected. If it’s negative, the hybrid does worse. We tested this on three small models — Falcon3-1B-Base, LLaMA-3.2-1B, and a quick check on Qwen2.5-1.5B — at different pruning levels and with both 4-bit and 8-bit quantization, all on a single consumer GPU (an RTX 3080 Ti). What we found: truthfulness up, reasoning down The results split sharply by task. On TruthfulQA, the hybrids consistently beat expectations — by about +7 points on Falcon and +12 to +14 points on LLaMA. It’s as if combining the two tricks accidentally made the models more careful about truthfulness. On reasoning and knowledge tasks, the opposite happened. GSM8K (grade-school math) and MMLU (general knowledge) dropped far more than expected — by 16 to 20 points on Falcon. In several hybrid setups, GSM8K accuracy fell all the way to zero. The twist: more pruning doesn’t mean more damage We assumed these effects would get worse the more we pruned. They didn’t. Instead, they peaked at moderate pruning (20–30%) and faded back toward normal at 50% pruning — both the TruthfulQA boost and the reasoning drop. More pruning didn’t mean more damage; it meant the surprise effect faded. We also found the order barely mattered — quantizing before pru…
AIhub · Research news and interviews · Original source

AI4Science Daily · 2026-09-24 — Claude discovers a novel enzyme system with CRISPR-like repeats

Mon, 05 Oct 2026 22:15:24 +0000

Selected developments

Claude discovers a novel enzyme system with CRISPR-like repeats
Science Sep 23, 2026 We’re introducing a new life sciences research group and laboratory at Anthropic. Our focus is on fundamental biology research using Claude: exploring datasets of DNA to identify uncharacterized protein families, generating hypotheses at scale, and testing them through experiments in the lab. This post introduces the team behind this work and shares early results in which Claude discovered a novel enzyme system with properties reminiscent of CRISPR, with only high-level direction from our scientists. Many discoveries that have revolutionized biology and medicine started with a scientist noticing something odd in the staggering diversity of molecular machines found in nature. Restriction enzymes, proteins that cut DNA at specific short sequences, were found in bacterial immune systems, where they destroy the DNA of invading viruses. Researchers realized they could use these enzymes to cut DNA at chosen places and splice genes from one organism into another, which launched the biotechnology industry. Taq polymerase, an enzyme that copies DNA at high temperatures, was identified in a bacterium in a Yellowstone hotspring. It became the basis for PCR, the DNA-copying method used in much of modern diagnostics. CRISPR was first noticed as an unusual repeat sequence in the DNA of certain bacteria, and is now the foundation of gene editing-based medicines. In the spring of 2026, we formed a research group to see whether general AI models can systematize and accelerate such discoveries. We believe that this acceleration will come from establishing a new way of doing biology research, in which agents collaborate with humans in every step of the process. Developing this new way of working required that we build our own lab and a single team working on everything from training Claude in biology to running experiments in the lab. Today, we’re sharing early results from one of our first research programs, in which Claude autonomously discovered a novel enzyme system that is associated with an array of DNA repeats, a pattern reminiscent of CRISPR. Although we don’t yet know its function, the system that Claude discovered has a set of characteristics that have only ever been found together in a handful of other systems, all of wh…
Anthropic · Science Blog · Original source
Anthropic says Claude found new enzyme system ART that may point to a new gene-editing mechanism
Anthropic touts AI-led biology discovery - Anthropic's AI model Claude discovered a new enzyme system called ART by analysing DNA sequences, potentially revealing a new gene-editing mechanism similar to CRISPR. - The AI autonomously analysed over 200,000 genes in 21 hours, accelerating research that usually takes weeks or months. - The discovery, still preliminary and unreviewed, could lead to advances in gene therapy and disease treatment, with experts calling for further investigation. AI generated SAN FRANCISCO – US artificial intelligence start-up Anthropic on Sept 23 unveiled the first discovery from its molecular biology laboratory, a previously unknown enzyme system spotted by its Claude model and touted by the company as proof that AI can speed up basic research. Enzyme systems found in nature have led to some of biology’s biggest breakthroughs, from the tools that launched the biotech industry to the DNA-copying method behind many medical tests. According to the company, some 950 Claude-powered software agents combed through a database of DNA sequences for 21 hours. The agents found more than 200,000 genes for one type of enzyme and picked out 20 worth a closer look – work that would take an expert scientist weeks or even months. Anthropic said its researchers’ involvement was limited to the initial instructions and the lab experiments, with the AI agents working autonomously to decide which leads to pursue. One of the agents spotted something in viruses that infect bacteria that scientists appear to have missed: a string of short, repeating pieces of DNA sitting next to two genes, one for a known enzyme and one for a mystery protein. The set-up looks like CRISPR, a natural defence system in bacteria that scientists turned into a gene-editing tool now used in medicine. Anthropic chief Dario Amodei, who recently called for slowing AI development because of its risks, cited the discovery on Sept 23 in remarks to the UN Security Council, which he addressed by video link. The finding is “preliminary”, he said, but “may constitute a new gene-editing mechanism that could have applications in gene therapy”. He predicted “profound progress in the near future on diseases that have plagued humanity for millennia”. The function of the newly identified system, dubbed ART, remains unknown, Anthropic said. The company added that its initial experiments suggest t…
Discovery · Anthropic · Original source
Tapio Schneider on understanding and trust in AI for mathematics, science, and engineering
[This is a guest post by Tapio Schneider . This blog post was initially written in a different file format and converted using AI. — T.] [This will be cross-posted on the CliMA blog .] The apparent proof of finite-time blow-up of the forced Navier-Stokes equation, announced by OpenAI on September 8, has brought into focus a debate about the role of AI in mathematics. The proof was produced with a system of some 10,000 AI agents that explored many approaches in parallel; it was then formalized and verified in Lean. The formal verification supports its correctness, but mathematicians are still working to digest it. A proof settles the truth value of a statement, but Terry Tao and others have argued that this is only part of the point of proofs; the other part is to advance our conceptual understanding of “ basic structures of shapes, numbers, and natural phenomena .” As Yehuda Rav put it a quarter-century ago, “ theorems are the headlines, proofs are the inside story .” A proof that is correct but incomprehensible, or undigested by the mathematical community, gives us the headline without the story. The issue is not whether the prover is a human or a machine, but whether the reasoning advances our collective understanding of methods and structures, which can spur new thinking and further advances. As Timothy Gowers has noted, the inadequate AI-generated write-ups of proofs are likely a temporary annoyance. I want to argue that the same point about understanding holds in the natural sciences and engineering, where it also has intrinsic value and, in addition, acquires instrumental value when predictions must be trusted before they can be verified empirically. One role of science is what ancient philosophers called episteme , roughly explanatory understanding; here, understanding is the goal itself. Another role is techne , roughly the craft of predicting and making: forecasting how natural or engineered systems will behave under circumstances not yet observed. Whether the two roles can be separated depends on how easily predictions can be checked. When they can, techne can stand on its own. A black-box AI weather prediction model can predict tomorrow’s weather, and we can trust it because its forecasts can be checked every day. Techne can then also serve episteme as an instrument. For example, AlphaFold predicts 3D protein structures without providing explanat…
Terence Tao · Original source
Anthropic Says Claude Spotted a Previously Unknown Enzyme System
Anthropic touts AI-led biology discovery Andrew Zinin Chief Editor U.S. artificial intelligence startup Anthropic on Wednesday unveiled the first discovery from its molecular biology lab, a previously unknown enzyme system spotted by its Claude model and touted by the company as proof that AI can speed up basic research. Enzyme systems found in nature have led to some of biology's biggest breakthroughs, from the tools that launched the biotech industry to the DNA-copying method behind many medical tests. According to the company, some 950 Claude-powered software agents combed through a database of DNA sequences for 21 hours. The agents found more than 200,000 genes for one type of enzyme and picked out 20 worth a closer look—work that would take an expert scientist weeks or even months. Anthropic said its researchers' involvement was limited to the initial instructions and the lab experiments, with the AI agents working autonomously to decide which leads to pursue. One of the agents spotted something in viruses that infect bacteria that scientists appear to have missed: a string of short, repeating pieces of DNA sitting next to two genes, one for a known enzyme and one for a mystery protein. The setup looks like CRISPR, a natural defense system in bacteria that scientists turned into a gene-editing tool now used in medicine. Anthropic chief Dario Amodei, who recently called for slowing AI development because of its risks, cited the discovery on Wednesday in remarks to the U.N. Security Council, which he addressed by video link. The finding is "preliminary," he said, but "may constitute a new gene editing mechanism that could have applications in gene therapy." He predicted "profound progress in the near future on diseases that have plagued humanity for millennia." The function of the newly identified system, dubbed ART, remains unknown, Anthropic said. The company added that its initial experiments suggest the system could, like CRISPR, be programmable. The results, published as a preprint, have not yet been reviewed by other scientists. Feng Zhang, one of the pioneers of CRISPR and a professor at MIT and the Broad Institute, called the identification of the repeated DNA fragments "genuinely intriguing" and said it "merits further investigation" in remarks released by Anthropic. Amodei also said on X that a team at Stanford University had independently desc…
Discovery · Anthropic · Original source
FutureHouse and Edison Scientific Assemble 12 Millennium Problems for Biology
Drawing inspiration from the Millennium Prize Problems in mathematics, Edison Scientific and FutureHouse have assembled a set of Millennium Problems for Biology. Each problem is intended to be very difficult to solve, but easy to validate in a simple laboratory environment. For this reason, grand challenges like “Solve Aging” that are very important and challenging but very hard to validate are excluded. Solving any one of these problems would represent a major advance in biology or biotechnology, and most of them would contribute materially towards curing disease. These are, in some sense, the “last reasonable eval” for AI in biology. The initial list of 12 problems was developed by Sam Rodriques and Michaela Hinks, with feedback from the FutureHouse and Edison Scientific teams. You can read more about each on the Millennium Problems for Biology webpage. - Origins of life - Cryopreservation - The reverse translatase - Improve Rubisco - The quadruplet cell - Somatic limb regeneration - Bacterial production of gene therapies - Programmable proteases - Cell-penetrating protein binders - Protein amplification chain reaction - 5' polymerases - New nitrogenases This is a living project maintained by the FutureHouse team and may be refined or expanded as science advances. We welcome your input – submit your comments on the current list of problems or propose new ones through our submission form. Stay tuned for further updates in the coming months as we iterate on the initial set. Questions can be directed to millenniumproblems@futurehouse.org.
FutureHouse · Original source

AI4Science Daily · 2026-09-23 — Bio-security Is an AI Arms Race: Radical Numerics CEO Eric Nguyen on Genomic Language Models

Wed, 07 Oct 2026 10:16:11 +0000

Selected developments

Bio-security Is an AI Arms Race: Radical Numerics CEO Eric Nguyen on Genomic Language Models
Radical Numerics co-founder Eric Nguyen discusses genomic language models and biosecurity. He helped develop Evo and Evo 2; a separate Arc/Stanford team later used those models to generate and synthesize functional bacteriophage genomes. The interview distinguishes three sequence scales: an average human gene is about 60K, long genes can reach 2.3M, and the whole human genome is around 3B. Nguyen argues that models which increase biological capability can also help defense keep pace. The 2.3M figure describes a long gene, not the whole genome.
Latent Space · Original source
XtalPi AI Enables Seven-Month PCC Nomination for Viva-Thera's MPER-Targeting HIV-1 Program
**media[1229066]**Viva-Thera makes its public debut with four modalities pursuing HIV prevention, treatment and cure, supported by XtalPi's platform capabilities.Viva-Thera combines its first atomic-resolution MPER-TMD-CT structure with XtalPi's AI and automation platform to advance a small-molecule PCC.MPER vaccine immunogens elicit antibody levels approximately five times those of the study comparator in rhesus macaques; antisera retain antiviral activity at a 1:1,000 dilution.BOSTON and SHENZHEN, China, Sept. 23, 2026 /PRNewswire/ -- XtalPi (HKEX: 2228) announced that Viva-Thera, a biotechnology company it incubated and backs as a strategic shareholder, has nominated a preclinical candidate (PCC) targeting the membrane-proximal external region (MPER) of HIV-1 gp41. The partners completed discovery and optimization in seven months by combining Viva-Thera's atomic-resolution structural research with XtalPi's AI drug design and automation platform. In its public debut, Viva-Thera also reported functional vaccine results in rhesus macaques and progress in targeted gene delivery. The company is pursuing HIV prevention and cure through four modalities-vaccines, long-acting small molecules, gene therapy and cell therapy-with XtalPi providing computational design, predictive modeling and experimental automation across the portfolio.
Discovery · XtalPi · Original source
John Platt on AI for Science: Google's ERA and the Scoreable Task
How often do you get to talk to a guest who has both an Academy Award and who invented textbook machine learning algorithms? John Platt has an Oscar , two textbook algorithms , two named asteroids, and an Erdos-Bacon number of 6. This was easily the most fun bio of all the guests we’ve read to date. And the result was an epic and fun chat covering Google’s Empirical Research Assistance (ERA), how AI can help battle climate change, and tons of great stories about the co-evolution of science and AI. John’s colleague Dave Bacon likes to tease John that his career has been defined by being twenty years early to the next big thing. This may be convolutional neural networks (some credit him with coining the term), fusion research, quantum computing. John and Google have been working on solving some of humanity’s hardest problems with AI and computation for well over a decade now. Recently John and his team set their sights on using AI to solve any scientific problem that can be written down as a score. Google’s Empirical Research Assistance (ERA) John’s team has taken on many hard scientific problems over the years. In solving these, they noticed a pattern, many scientific problems can be reduced to what John calls a “scoreable task”. Once you have the score function, the goal is to find some code that maximizes the score. The hard part is in formulating the score, but once you have the score finding the maximizer can still be quite a lot of effort. John’s team set out to automate solutions to this general problem. This came out of the idea of an “auto-Kaggle” AI, which can solve any Kaggle problem you can throw at it. Kaggle is owned by Google, so all the data was ready and easily available to them! The result is Google’s Empirical Research Assistance or ERA ( paper , github , blog ). 1 ERA is surprisingly simple conceptually. Gemini (or your LLM of choice) keeps a running tree of past experiments (notebooks) and where they’re going. It’s a close cousin of Monte Carlo Tree Search : at each iteration the Upper Confidence Bound rule picks which notebooks are most promising to mutate. This is optimistic, not greedy, so sometimes even the fifth-best notebook gets chosen. Gemini then proposes mutations for each one, about ten at a time. The history of each branch is shared, so different leaves can learn from each other. “It’s almost like having a hyper-eager grad stu…
Latent Space · Original source
Guiding large language models to predict edit sequences for molecular synthesizability optimization
Nature Machine Intelligence, Published online: 23 September 2026; doi:10.1038/s42256-026-01304-x Computational drug design often generates molecules that cannot be synthesized. Li and Lai show how large language models can predict precise structural edits to fix these molecules, outperforming traditional methods and preserving key features.
Nature Portfolio · Nature Machine Intelligence · Original source

AI4Science Daily · 2026-09-22 — Anthropic’s Biology Lab: Secondary Robotics Report and Official Lab-Work Clarification

Thu, 24 Sep 2026 03:51:14 +0000

Selected developments

Anthropic’s Biology Lab: Secondary Robotics Report and Official Lab-Work Clarification
This September 22 secondary report describes Anthropic’s Bay Area biology laboratory and reports plans for Claude-directed robotic experiments. In its September 23 announcement, Anthropic explicitly states that all lab work in the laboratory it describes is performed by human scientists. Claude searches genomic data, proposes candidate systems and assists analysis. The two sources differ in date and claim; the earlier report does not establish an operating autonomous robotic laboratory. The function of the ART enzyme system described in the official announcement remains under investigation.
Discovery · Anthropic · Original source
First-in-Human Phase I Data for Insilico's AI-Designed Pan-TEAD Inhibitor ISM6331 to Be Presented at ESMO 2026
Rapid Oral presentation will report initial Phase I data for ISM6331 in patients with advanced solid tumors FDA Fast Track- and Orphan Drug-designated program reflects Insilico’s strategy to advance AI-designed medicines against difficult-to-treat, age-associated diseases CAMBRIDGE, Mass. — September 22, 2026 — Insilico Medicine ("Insilico"; HKEX: 3696), a clinical-stage generative artificial intelligence (AI)-driven drug discovery company, today reminded the oncology community that initial Phase I data for ISM6331, its investigational pan-TEAD inhibitor, will be presented in a Rapid Oral session at the European Society for Medical Oncology (ESMO) Congress 2026, taking place October 23–27 in Madrid, Spain. The presentation will provide the first reported clinical findings from the global, multicenter first-in-human study of ISM6331 in patients with advanced solid tumors. ISM6331 is designed to inhibit all four TEAD family transcription factors and disrupt tumor-promoting signals driven by dysregulation of the Hippo pathway, a central regulator of cell growth, survival, tissue homeostasis and regeneration. Presentation details Title: First-in-human Multicenter Phase 1 Study of ISM6331, an AI-Designed Pan-TEAD Inhibitor, in Patients with Mesothelioma or Other Advanced Solid Tumors Abstract number: #997 Session: Rapid Oral Presentation: Developmental Therapeutic Format: Rapid Oral presentation Presenter: Melina E. Marmarelis, MD, Hematology and Oncology, Abramson Cancer Center, University of Pennsylvania Date: Sunday, October 25, 2026 Time: 8:30–10:00 a.m. CEST Location: Cordoba Auditorium – Hall 4 “The ESMO presentation is an important opportunity to share the first clinical findings from ISM6331 with the global oncology community,” said Halle Zhang, PhD, Vice President, Clinical Development – Oncology at Insilico Medicine. “Patients with advanced mesothelioma and other Hippo pathway-dysregulated tumors have a significant unmet medical need. We look forward to presenting the emerging safety, pharmacokinetic and antitumor findings from this first-in-human study and discussing the next stage of the program.” From generative chemistry to clinical validation Insilico nominated ISM6331 as a preclinical candidate in June 2023. Its novel, non-covalent scaffold was generated with support from Chemistry42, the company’s generative chemistry platform, using structure-b…
Discovery · Insilico Medicine · Original source
Intuitive enzyme design with LLM agents
A collaborative system of large language model agents brings protein engineering closer to natural-language interaction, lowering the barrier between human intent and biomolecular design. This is a preview of subscription content, access via your institution Access options Access Nature and 54 other Nature Portfolio journals Get Nature+, our best-value online-access subscription 27,99 € / 30 days cancel any time Subscribe to this journal Receive 12 digital issues and online access to articles 111,21 € per year only 9,27 € per issue Rent or buy this article Prices vary by article type from$1.95 to$39.95 Prices may be subject to local taxes which are calculated during checkout References - Shao, Q. et al. Nat. Comput. Sci. https://doi.org/10.1038/s43588-026-01049-y (2026). - Wiesinger, J., Marlow, P. & Vuskovic, V. Agents. Kaggle https://www.kaggle.com/whitepaper-agents (2025). - Ma, Z. et al. J. Chem. Inf. Model. 65, 6599–6612 (2025). - Ferruz, N., Schmidt, S. & Höcker, B. Nat. Commun. 13, 4348 (2022). - Nijkamp, E., Ruffolo, J. A., Weinstein, E. N., Naik, N. & Madani, A. Cell Syst. 14, 968–978.E3 (2023). - Ghafarollahi, A. & Buehler, M. J. Digit. Discov. 3, 1389–1409 (2024). - Ponnapati, M. et al. ProteinCrow: a language model agent that can design proteins. In Generative AI for Biology Workshop, Int. Conf. Machine Learning https://icml.cc/virtual/2025/51218 (ICML, 2025). - Durumeric, A. E. P. et al. J. Chem. Inf. Model. 66, 5721–5735 (2026). - Dolorfino, M. et al. Assessing the generalizability of machine learning and physics-based methods with DNA-encoded libraries. Preprint at bioRxiv https://doi.org/10.64898/2026.04.18.719394 (2026). Ethics declarations Competing interests The author declares no competing interests. Rights and permissions About this article Cite this article Sztain, T. Intuitive enzyme design with LLM agents. Nat Comput Sci (2026). https://doi.org/10.1038/s43588-026-01052-3 - Published: - Version of record: - DOI: https://doi.org/10.1038/s43588-026-01052-3
Nature Portfolio · Nature Computational Science · Original source
Computational active materials and embodied intelligence in extreme conditions
Nature Computational Science, Published online: 22 September 2026; doi:10.1038/s43588-026-01051-4 Embodied intelligence has enabled machines to act autonomously, yet intelligence remains largely confined to digital models, leaving robotic bodies as passive executors. In extreme environments, such as the human brain, where tissues are mechanically and thermally sensitive, conventional electronics can become limiting. Here, we discuss how active materials, which embed intelligence directly within their structure, can expand robotic capabilities in these settings, with a particular focus on the role of computation-aided design.
Nature Portfolio · Nature Computational Science · Original source
Did OpenAI Solve the Wrong Navier-Stokes Problem?
Two weeks ago OpenAI claimed a solution to one of the biggest open problems in math—the Navier-Stokes problem—an achievement worth a $1-million prize from the Clay Mathematics Institute. The proof ignited a powder keg of concern over artificial intelligence companies’ race to disrupt the subject. But with the dust still far from settled, a different controversy is emerging: Did OpenAI even solve the right Navier-Stokes problem? Generated by an internal large language model (LLM), OpenAI’s proof relies on an approach that many experts find unnatural. It solves a variant of the problem that mathematicians say is disconnected from reality and thus less interesting. In a sense, the LLM found and exploited a loophole in the framing of the question. On supporting science journalism If you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today. “The most important problem is unsolved,” says Luis Silvestre, a mathematician at the University of Chicago. “The Clay problem is settled, but the main problem for the Navier-Stokes equations is not.” Furthermore, last Thursday, three mathematicians posted a proof of their own that showed that OpenAI’s method can never be extended to solve the full problem. In other words, the loophole will never be closed, barring some completely new idea. The Navier-Stokes equations are supposed to describe how fluids flow, but mathematicians doubt whether they can always be trusted. The million-dollar problem is about whether the equations ever “blow up,” which would mean they’d allow the flow to be infinitely fast at points—something that can’t happen in the real world. But there’s one piece of the equations that’s optional—sometimes it’s there; sometimes it isn’t. The “it” here is an external force such as gravity that affects how a fluid moves. “All fluids we know of are under some kind of external force,” says mathematician Diego Córdoba. “So to have the force makes complete sense.” When most experts think about the Navier-Stokes problem, though, it’s without this force. They want to find a purer, more fundamental way for the equations to blow up by using only the intrinsic forces within any fluid—not by applying some specific, precisely contrived external f…
Discovery · OpenAI · Original source

AI4Science Daily · 2026-09-21 — EdgeGen: Improving Tool-Calling Agents Beyond Happy Paths with Synthetic Edge Case Generation

Fri, 09 Oct 2026 10:11:40 +0000

Selected developments

EdgeGen: Improving Tool-Calling Agents Beyond Happy Paths with Synthetic Edge Case Generation
Tool-calling LLM agents are increasingly deployed in enterprise applications. However, effective evaluation and optimization require high-quality, diverse task datasets that are often difficult to obtain due to privacy and other constraints. Existing synthetic task generation methods often produce generic tasks that ignore an agent's underlying state or database and fail to reflect real-world usage diversity. We propose EdgeGen, a synthetic task generation framework that extracts compliance rules from an agent's specification and uses them to generate database-grounded edge-case tasks designed to violate these rules. When combined with existing synthetic data generation techniques, EdgeGen enables agent improvement through finetuning and harness optimization. The resulting pipeline forms a fully automated closed-loop system that requires no human annotation. Finetuning on data generated by EdgeGen yields a consistent mean progress improvement of 2 percent to 42 percent on tau2bench airline domain, while other baseline methods show degradation for some models. On the other hand, for harness optimization, our method shows a mean progress improvement of 10 percent and 30 percent over the human-curated and base harnesses, respectively, for the Gemma-4-e4b model.
Hugging Face Daily Papers · Original source
Emergent Collusion in Long-Horizon LLM Agent Interaction
LLM agents are increasingly deployed in collaborative settings, yet long-term interaction may give rise to undesirable coordination. We study the emergence of collusion in a long-horizon multi-agent environment: two agents repeatedly complete individual tasks, share task logs, verify each other's work, and receive rewards. We introduce realistic constraints that make compliance with the verification protocol incompatible with reward maximization, and find that agents increasingly deviate from the protocol over repeated interactions. Collusion emerges in 94% of trajectories across 10 models, and more capable models within the same family reach it earlier. Controlled peer interventions show that collusion is shaped by peer behavior, while ablations reveal additional effects of reward structure, the verification feedback agents receive, and their interaction history. In particular, restricting the amount and scope of interaction history available to agents reduces collusion. Overall, our findings show that long-horizon interaction can reshape how agents coordinate in ways that create safety risks.
Hugging Face Daily Papers · Original source

AI4Science Daily · 2026-09-18 — How Claude is uplifting biomolecular modeling

Fri, 18 Sep 2026 07:26:23 +0000

Selected developments

How Claude is uplifting biomolecular modeling
Claude made the open-source models that scientists use to predict and design biomolecules faster and more memory-efficient. Claude optimized more than 30 of these models in just under four weeks, speeding them up roughly 4x on average. It also created a low-memory mode that enables the accurate prediction of biomolecular systems larger than 10,000 tokens (amino acids, nucleotides, and atoms from small molecules and ions) on a single NVIDIA GPU node.
Anthropic · Science Blog · Original source
ScienceIDE: Turning the World's Scientific Codebase into Agent-Learnable Environments
ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments 摘要 Scientific code repositories encode decades of human knowledge in executable models, methods, and tools. Yet fragmented toolchains, implicit domain conventions, and specialized correctness criteria make this knowledge difficult to convert into reliable learning experience—a challenge we call the scientific experience bottleneck. We introduce ScienceIDE, infrastructure for turning the world’s scientific code into programmable environments for scientific agents. Guided by expert-defined scientific cases and acceptance criteria, agents transform repositories into executable environments that support task generation, execution, and scientific verification. These environments provide a shared foundation for supervised fine-tuning, reinforcement learning, and evaluation. Using verified interaction trajectories, we train PhAI-IDE-72B, PhAI-IDE-9B, and PhAI-IDE-4B. The model family shows gains in held-out scientific-code repair and across selected general-purpose benchmarks in code, reasoning, and knowledge, providing evidence of positive transfer from scientific experience to broader capabilities. ScienceIDE lays the foundation for an integrated workspace for agent learning and scientific practice, making humanity’s scientific software a shared substrate for developing scientific intelligence. 作者 * 共同一作期刊/会议引用 BibTeX @article{geng2026scienceide, title={ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments}, author={Geng, Hejia and Huang, Zesen and Li, Haoyang and Li, Wenbin and Wu, Koutian and Zhou, Zihan and Pang, Yuanbo and Liu, Weihao and Xu, Zigong and Li, Zhiping and Zhang, Zongzheng and Dong, Chuanfei and Sun, Jiankai and Zheng, Tianzhe and Xie, Fengyu and Ma, Yue and Shi, Yueheng and Wang, Junkai and Xie, Tong and Di, Zonglin and Liu, Xianrong and Gao, Qucheng and Liu, Yimin and Pan, Jiaming and Huang, Sheng and Ma, Xiao-Han and Yuan, Lanqing and Zhu, Zhenlin and Liu, Ziang and Xu, Ziyang and Liang, Kangkai and Xian, Jiayi and Zhao, Zehong and Ma, Taigao and Chang, Fuming and Hu, Yuanzhe and Xu, Liuwei and Tang, Xianzhe and Xie, Jingxu and Zhang, Peijin and Gao, Qiang and Xing, Chengyi and Zhao, Zhe and Wang, Xi and Wang, Leo and Wu, Fang and Xing, Yaopeng and Meng, Xing and Yin, Zhenfei and Wu, Yingcheng and Yang, Ling}, journal={arXiv preprin…
PhAI Labs · Original source
Changing minds: How AI is transforming the life sciences
Artificial intelligence (AI) is changing the way biomedical research is conducted by making it possible to integrate diverse types of biological data and tackle questions that were previously difficult to address. At the same time, its growing use also raises important questions about data quality, reliability, interpretability, and how computational predictions should be combined with biological knowledge and experimental validation. In this Voices piece, researchers from diverse disciplines share their perspectives on how AI is advancing their respective fields and highlight both the transformative opportunities and the key challenges that will shape the future of the life sciences.
Cell Press · Cell · Original source
What a crowdsourced game revealed about steering Olmo 3
AI models need to do more than produce correct answers. How they respond matters too: whether they’re helpful, fair, safe, respectful, and responsive to the people using them. For model builders, the challenge is knowing whether those “prosocial” behaviors hold up in practice—and whether evaluations capture how a model behaves when people interact with it in unexpected ways. Northeastern University MS student Soham Padia used Olmo 3 to test whether crowdsourcing an evaluation of prosocial behavior could expose weaknesses that a small research team might miss. Padia had developed an evaluation that measures how strongly text steers a model toward more prosocial responses. Steering Arena turned that evaluation into a sort of game—players submit short text prefixes designed to influence the model, see how strongly each one shifts Olmo 3 in that direction, and compete for the top spot on the leaderboard. Olmo’s openness made the project possible—Padia could see how submitted text changed Olmo 3’s internal activity instead of inferring those effects only from the responses it generated. That access became the foundation for both his evaluation and Steering Arena. From open access to a public challenge Padia chose Olmo 3-32B so he could study prosocial steering in a relatively large model. Through the National Deep Inference Fabric (NDIF), a U.S. National Science Foundation (NSF)-supported platform for experimenting with large open models, he could access Olmo 3-32B remotely without owning the GPUs needed to host it himself. That effort to make advanced AI research more accessible aligns with Ai2’s work with NSF. Through the OMAI project, Ai2 is developing fully open models and infrastructure designed to help more researchers study, reproduce, and build on sophisticated AI systems. "Open weights alone would not have been enough," Padia says. "Olmo documents its data and its post-training, so when I find a prosocial direction inside it I know whether I am looking at something the pretraining put there or something a later fine-tune installed. On most models, that question simply has no answer." Padia’s evaluation uses 135 pairs of contrasting text responses spanning 15 qualities, including empathy, fairness, safety, privacy, and respect. (Each pair starts with the same prompt and contrasts a more prosocial response with a less prosocial one.) By comparing the mode…
Allen Institute for AI · AI for Science · Original source
Multi-organelle signatures map cell-state diversity and metabolic adaptation in tissues
Cell-state diversity drives tissue adaptability, repair, and disease resilience, but capturing this complexity is a challenge. Current approaches rely on transcriptional profiling and overlook organelle structure, a key indicator of metabolism and stress. We developed spatial Organellomics (sOrganellomics), an imaging workflow that integrates automated segmentation with machine learning to classify and spatially map cell states from multi-organelle signatures. In liver and pancreas, these signatures distinguished broad cellular classes. In liver, sOrganellomics revealed that zonal position did not fully explain organelle-defined hepatocyte categories. Instead, hepatocytes formed intermixed communities within canonical zones, supporting a refined subzonal diversity model. Nutritional stress reshaped this organization. Intravital imaging linked fasting-induced organelle remodeling with altered mitochondrial membrane potential in vivo, supporting multi-organelle architecture as a structural readout of tissue adaptation.
Science · Science Current · Original source

AI4Science Daily · 2026-09-17 — An expert-level generalist AI for abdominal CT diagnosis

Thu, 24 Sep 2026 10:19:29 +0000

Selected developments

An expert-level generalist AI for abdominal CT diagnosis
Artificial intelligence (AI) in radiology aspires to deliver expert-level diagnosis across diverse clinical tasks, yet existing supervised strategies remain limited in scope. We developed RADAR, a generalist vision-language model trained on more than 400,000 contrast-enhanced abdominal computed tomography (CT) examinations and 15 million anatomy-wise image-text pairs, learning directly from clinical reports without manual annotation. Throughout internal and external evaluations across multiple centers and varied clinical scenarios, RADAR achieved high diagnostic performance and robust generalization for 18 anatomical structures and 146 imaging findings. In a reader study, RADAR assistance increased the diagnostic sensitivity of 26 radiologists by ~10%. RADAR offers a scalable, versatile, and interpretable solution for abdominal CT, demonstrating that generalist AI can match human experts in general and complicated radiology tasks.
Science · Science Current · Original source
scBaseCount: An AI agent-curated, standardized, auto-updated single-cell data repository
scBaseCount is presently the largest public single-cell RNA-seq repository, containing over 502 million cells across 27 organisms and 75 tissues. An AI agent autonomously discovers, annotates, and uniformly reprocesses all 10× Genomics datasets in the SRA, creating a harmonized, continually updated resource for studying the diversity of cell biology and training AI models.
Cell Press · Cell · Original source
How fast are you ageing? Ask AI
Understanding why people age — and how to slow that process down — is an obsession that has stretched across millennia, from mythic quests for immortality to modern science’s hunt for drugs that lengthen lifespan. Now, that quest has entered the AI era. A report1 published today in Cell unveils a bespoke artificial-intelligence system that aims to turbocharge longevity research. The system introduces large language models (LLMs) that the authors trained on ageing-biology data. It also includes a suite of 17 tasks that can serve as benchmarks to gauge how well these and other LLMs perform on ageing-related projects. Finally, it includes an interface that brings the models and other ageing-related research tools together with AI assistants called agents to aid analyses. Gene clock predicts time to death in humans — and assesses ‘biological’ age The authors used their benchmarks to evaluate both their specialist LLMs and much larger cutting-edge commercial models produced by companies including OpenAI and DeepSeek, that are trained on larger, more diverse data sets to handle a broad range of tasks. In many — although not all — of the tests, the LLMs tailored to ageing research outperformed the large LLMs. “The paper makes a very important contribution to the longevity and ageing field,” says Marinka Zitnik, a computer scientist at Harvard Medical School in Boston, Massachusetts, who was not involved in the work. “It can inform the development of next-generation AI models.” The work also tackles a key question for the field: how can researchers train AI tools to understand ageing when scientists have not yet defined the concept for themselves? “For certain disease areas, such as cancer, the AI tasks can be very crisply defined,” says Zitnik. “Ageing is a very different beast. It is hard to define.” Clock collectors Researchers have spent decades gathering data to define that beast. One result has been a slew of ‘clocks’ aimed at measuring biological ageing, or the rate at which a person’s body exhibits signs of ageing. Such clocks measure biological age by assessing characteristics such as facial features, protein levels, gene activity and brain scans, all with the goal of quantifying whether the effects of ageing are outpacing or lagging behind someone’s chronological age in years. How quickly are you ageing? What molecular ‘clocks’ can tell you about your hea…
Nature Portfolio · Nature · Original source
Revisiting molecular descriptors with TDiMS for interpretable intramolecular interactions based on substructure pairs
Abstract Molecular descriptors play a crucial role in representing the structural features of molecules for machine learning-based physical property prediction. However, current descriptors either consider only local aspects of molecular structures or fail to effectively learn nonlocal structural features involving long-distance intramolecular interactions. Here, to address this issue, we present a descriptor named TDiMS. TDiMS effectively summarizes the enumerated pairwise topological distances between molecular substructures, thus capturing nonlocal interactions. Our evaluation shows that TDiMS successfully identifies essential features of large structures and outperforms other representative descriptors in predicting properties for which distances between substructures are a primary factor. In addition, these identified features are highly interpretable for experts in materials discovery.
Nature Portfolio · Nature Computational Science · Original source
JEPA-Anything: Learning Predictive Models across Different Worlds
JEPA-Anything: Learning Predictive Models across Different Worlds 摘要 World modeling enables intelligence to anticipate consequences, guide interventions, and learn from interaction. Yet predictive models remain domain-specific: can a common learning principle support world modeling across radically different systems? We introduce JEPA-Anything, a domain-agnostic framework based on orthogonal predictive factorization (OPF). Extending joint-embedding predictive architectures, OPF decomposes latent targets into complementary factors, learns them through dedicated pathways, and recombines them within a shared predictive design. We evaluate JEPA-Anything across seven domains: vision, biology, clinical trajectories, control, molecular dynamics, physical fields, and weather. Experiments span representation learning, intervention prediction, out-of-distribution generalization, and long-horizon dynamics, including 10 matched dynamics tasks, forecasting of over 1,000 clinical events, and 100-step molecular rollouts across four systems. Against matched JEPA baselines, JEPA-Anything improves reported metrics on all 10 dynamics tasks and reduces single-intervention prediction error on Interventional Pong by 34.8%. It achieves the lowest one-step and 100-step molecular errors among compared methods in all four systems. Beyond prediction, a factor-nominated biological intervention receives experimental support in cell co-cultures, patient-derived organoids, tumor fragments, and mice; latent orbital modes recover the Keplerian scaling exponent with a fitted slope of −1.4991. These results support a common factorized predictive principle across heterogeneous worlds, connecting world modeling with intervention and experimentally grounded scientific discovery. 作者期刊/会议引用 BibTeX @article{cui2026jepa, title={JEPA-Anything: Learning Predictive Models across Different Worlds}, author={Cui, Taoyong and Wang, Zhongyao and Xu, Xinyue and Liu, Weiyang and Yu, Zhaochen and Zhang, Yuying and Gao, Qiang and Yang, Mengyue and Ouyang, Wanli and Heng, Pheng Ann and Wu, Yingcheng and Yin, Zhenfei and Yang, Ling}, journal={arXiv preprint arXiv:2609.20800}, year={2026} }
Hugging Face Daily Papers · Original source

AI4Science Daily · 2026-09-16 — Discovery Foundation Models: Toward Open-Ended Discovery Intelligence

Wed, 16 Sep 2026 09:49:43 +0000

Selected developments

Discovery Foundation Models: Toward Open-Ended Discovery Intelligence
Discovery Foundation Models: Toward Open-Ended Discovery Intelligence 摘要 Foundation models have progressed from learning and reasoning over existing knowledge, to increasingly learning through action, tool use, and outcome feedback. We argue that the next frontier is a further transition: from solving and acting within problems specified by humans to participating in the process by which new problems, representations, explanations, and knowledge are created. We refer to this capability as Discovery Intelligence. We formulate Discovery Foundation Models (DFMs) as general-purpose model systems for open-ended discovery. A DFM operates over a revisable research state and supports seven coupled capabilities spanning problem discovery, formulation, representation construction, hypothesis formation, intervention, evidence-grounded revision, and continual discovery improvement. We instantiate this framework with Zetema, which couples explicit research-state dynamics, verification and experimental gating, external grounding, and cross-task Discovery Skill evolution. We further ground the framework with GALILEO, a real therapeutic-discovery system in which Dry-Lab reasoning, robotic and hands-on Wet-Lab experimentation, external biological evidence, and iterative hypothesis and design revision form a closed physical discovery loop. We then formulate a unified approach to capability formation and process-centered evaluation, enabling discovery behavior to be trained, improved, and measured beyond final-answer performance. Together, these components establish discovery as a learnable, executable, and evaluable capability of foundation-model systems. We view this shift as a broader progression in intelligence scaling: from learning over existing knowledge, to learning from action outcomes, and ultimately to participating in the construction, testing, and revision of the structures through which new knowledge is discovered. 作者期刊/会议引用 BibTeX @article{yang2026discovery, title={Discovery Foundation Models: Toward Open-Ended Discovery Intelligence}, author={Yang, Ling and Yin, Zhenfei and Wu, Yingcheng}, journal={arXiv preprint arXiv:2609.15973}, year={2026} }
PhAI Labs · Original source
PhAI Labs introduces the DFM research framework
PhAI Labs announced its DFM research framework, published the DFM technical report and research-map repository, and said ScienceBuddy, ScienceIDE and JEPA-Anything were scheduled as separate releases on September 16–18.
PhAI Labs · Original source
LabAgent: Customize Any Research Hubs for Scientific Discoveries Using AI Agents
Scientific research is a continuous process that emphasizes inheritance. Methods developed by predecessors are often expanded upon by new researchers to explore more novel and in-depth scientific questions. However, the change of lab staff, such as student graduation, leads to a lack of personnel capable of replicating methods. Methods that have been developed with significant effort and resources cannot be continued. To address these limitations, we propose LabAgent, a reproduce and discovery harness tailored for a lab's continuous work. LabAgent employs two mechanisms to guarantee that all skills can be executed and verified and to record the corrective methods and experiences, allowing for direct correction or avoidance of similar errors. We applied LabAgent to drug property prediction, biomedical problem analysis, protein variant effect prediction, and statistical genetics in life science domains. LabAgent ranks first over commercial generalist agents in every domain, and demonstrates accurate reproduction of a published figure. Overall, these results demonstrate that LabAgent can effectively integrate and reasonably expand laboratory knowledge.
arXiv cs.AI · Original source
RAGCell: Retrieval-Augmented Generation as Supervision for Versatile Single-cell Analysis
Single-cell foundation models (scFMs) are transforming computational biology by enabling generalizable, task-agnostic representations for versatile single-cell analysis. Despite their progress in facilitating rapid deployment for downstream tasks, off-the-shelf scFMs still have some overlooked concerns: (I) (Pretraining Cost.) Pretrain-based scFMs necessitate pretraining on a vast volume of cells, rendering it draining resources in applications. (II) (Heterogeneous Gap.) Large Language Models (LLM)-based scFMs ignore the tremendous heterogeneous gap between LLM textual and raw cellular spaces, leading to insufficient capability when facing downstream tasks. To this end, we introduce RAGCell, a versatile single-cell analysis framework that achieves a double-win in both cost-effectiveness and high performance. The success of RAGCell lies in two key aspects: Leveraging LLMs to construct cell-level and feature-level knowledge databases, which serve as supervision signals for training the cell model and significantly reduce the training cost ($>$pretrain-based scFMs). Aligning cell representations with text embeddings from the bi-level knowledge databases, enabling knowledge transfer from textual spaces to cellular spaces and effectively mitigating the heterogeneous gap ($>$LLM-based scFMs). Through extensive experiments on six downstream single-cell analysis tasks, we demonstrate that RAGCell achieves outstanding performance compared to state-of-the-art scFMs while operating at less than $\sim$1/10 the cost of pretrain-based scFMs.
arXiv Genomics · Original source
Interview with Yash Saxena: how is external knowledge used in AI systems?
In a new series of interviews, we’re meeting some of the PhD students that were selected to take part in the Doctoral Consortium at the International Joint Conference on Artificial Intelligence and the 29th European Conference on Artificial Intelligence (IJACI-ECAI 2026) . In this first article of the series, we caught up with Yash Saxena to find out how he is trying to make it easier to understand how external knowledge is used in AI systems. Tell us a bit about your PhD – where are you studying, and what is the topic of your research? I am a PhD student in Computer Science at the University of Maryland, Baltimore County (UMBC), advised by Dr. Manas Gaur in the Knowledge Infused AI and Inference (KAI2) Lab. My PhD focuses on making AI systems that use external information easier to understand and audit. Many AI systems now look through external documents before answering a question. My research follows what happens to that evidence as it moves through the system. I study where it came from, why certain evidence was selected, whether it actually supports the claim being made, and whether the context needed to interpret that evidence correctly is preserved along the way. I call this broader research direction IKnowFlow , or Interpretable Knowledge Flow . The aim is to keep the evidence trail visible so that someone can inspect a generated claim and understand how the system arrived at it. This is especially important in areas such as law, science, and healthcare, where an incorrect AI-generated claim can have serious real-world consequences. Could you give us an overview of the research you’ve carried out so far during your PhD? I started by studying source attribution, or how an AI-generated claim can be linked back to the source that supports it. In REASONS (REtrieval and Automated citationS Of scieNtific Sentences), we evaluated retrieval and citation generation across more than 20,000 scientific papers from 12 research areas. I later compared two ways of adding citations to large language model responses: generating them together with the answer or assigning them afterward. In our human evaluation, adding citations afterward was more accurate, but this also showed that a citation alone does not necessarily reveal how the model arrived at its answer. From there, I started looking earlier in the process. Interpretable Modular Retrieval Neural Networks (IMR…
AIhub · Research news and interviews · Original source

AI4Science Daily · 2026-09-15 — Drug firms' secret data supercharge AI protein models

Wed, 23 Sep 2026 22:19:50 +0000

Selected developments

Drug firms' secret data supercharge AI protein models
For drug discovery, protein-folding models such as AlphaFold have a data problem: there isn’t enough of it in public databases. Improving the performance of these artificial-intelligence-based tools will require extra data that provide examples of how proteins and drugs interact, some scientists argue. Protein structures— locked away by the thousands in drug company vaults— offer one promising source. Today, a consortium of pharmaceutical companies reports that using such data to train AI models of protein folding improves model performance markedly. The group used OpenFold3 — an open-source replication of AlphaFold 3 — to develop a new model trained on more than 20,000 proprietary protein structures. The system outperformed both comparable ones trained on public data alone and those trained on the siloed datasets of individual firms. The study, described in a blog post, has not been peer-reviewed, and the model is not publicly available. “You add all this data, and you get a pretty big bump in performance,” says Mohammed AlQuraishi, a computational biologist at Columbia University in New York City, who was part of the effort. AlphaFold is running out of data — so drug firms are building their own version The findings, he says, strengthen the case for generating similar publicly available datasets to supercharge protein-folding AIs. One such project, called OpenBind and supported by up to £8 million (US$10.8 million) in UK government funding, released hundreds of new protein structures last month, with thousands more in the works. An untapped vein The Protein Data Bank (PDB), an open repository of more than 200,000 experimentally determined protein structures, was the bedrock of AlphaFold 2’s training data. It enabled the tool to predict protein structures with startling accuracy — a breakthrough recognized with the 2024 Nobel Prize in Chemistry. The model’s successors, including AlphaFold 3, added the ability to predict how proteins will interact with other molecules, including potential drugs. But the PDB has relatively few examples of experimentally determined structures interacting with drug-like molecules — maybe just 10,000, says Paul Mortenson, vice-president for computational chemistry and informatics at Astex Pharmaceuticals in Cambridge, UK. That lack of data is a problem for drug-discovery efforts. Research has suggested that the accuracy of Alph…
Nature Portfolio · Nature · Original source
An engineered nanopore identifies saccharides, amino acids, peptides and ribonucleotides
Simultaneous characterization of proteins, RNAs and glycans with a nanopore remains a substantial challenge. To advance toward this goal, we introduce a multifunctional adaptor, maleimido-C2-formylphenylboronic acid (maleimido-C2-FPBA), into the Mycobacterium smegmatis porin A (MspA) nanopore at the pore constriction site. Here we demonstrate that the engineered nanopore, MspA-FPBA, enables the simultaneous identification of a diverse set of analytes, including 21 proteinogenic amino acids, three post-translationally modified (PTM) amino acids, four canonical nucleoside monophosphates (NMPs), three epigenetically modified NMPs, four monosaccharides and five peptides. Combined with machine learning, this sensor achieves an overall accuracy of 98.7%. Furthermore, it generates predictable event signatures for specific analyte types, enabling the machine learning model to identify analytes without strictly relying on prearchived event features, as demonstrated by the nanopore analysis of yeast cell extract. Lastly, we apply MspA-FPBA for compositional analysis of G2 glycopeptides. Future integration of hydrolase with MspA-FPBA may further improve MspA-FPBA utility as a single system to identify diverse biomolecules. Modified nanopore simultaneously identifies diverse biomolecules and their modifications.
Nature Portfolio · Nature Biotechnology · Original source
A multi-period dataset of retrogressive thaw slump susceptibility for Northern Hemisphere permafrost regions
Permafrost thaw accelerates the initiation and expansion of retrogressive thaw slumps (RTSs), with critical implications for the terrestrial carbon cycle and land-surface energy budgets. However, long-term datasets characterizing spatiotemporal variations in RTS susceptibility across Northern Hemisphere permafrost regions remain lacking. Here, we integrate 55,673 georeferenced RTS occurrence records with machine learning models to generate a 1-km spatiotemporal dataset of RTS susceptibility from 2000 to 2100 under multiple climate change scenarios. RTS susceptibility is defined as the relative occurrence probability of RTS under specified environmental conditions. We independently modelled RTS susceptibility for the Arctic and Tibetan Plateau and merged the individual outputs to construct a spatially harmonized dataset for the Northern Hemisphere. Model performance was evaluated using standard classification metrics, yielding mean area under the receiver operating characteristic curve (AUC) values of 0.96 ± 0.005. The dataset includes georeferenced RTS inventories and gridded susceptibility maps for the entire Northern Hemisphere permafrost domain, covering the baseline period and diverse future climate scenarios. These products provide a foundational resource for quantifying permafrost hazard risks, infrastructure exposure, and carbon-climate feedbacks under climate warming.
Nature Portfolio · Scientific Data · Original source
Why Don't Machine Learning Research Agents Overfit?
Machine learning, at its core, is about generalization, not memorization. You hand your learning algorithm a pile of training examples and use them to fit a model. But the goal is not to perform well on the training examples — that's easy, you could just memorize the answers. The goal is to perform well on new examples that you have never before seen. If a model does well on the data it was trained on but poorly on fresh data, it hasn’t actually learned anything; you have only fooled yourself into thinking it has. This failure mode has a name: overfitting. Anyone who has taken an introductory statistics or machine learning class knows the standard defense. You hold out some of your data and refuse to train on it. In practice, this held-out data plays two roles. A validation set is one you consult repeatedly while building the model — to compare candidates, tune hyperparameters, and decide what to try next. A final test set (or holdout) is meant to be touched only once, at the very end: because the training procedure never saw it, strong performance there is a correct proxy for the new examples you will encounter in the wild. Machine learning, at its core, is about generalization, not memorization The “holdout” condition is crucial, though. The correct-proxy guarantee holds if the held-out set stays genuinely unseen. If you check your performance on it, tweak your training procedure in response, recheck, and iterate, chasing better and better numbers, that set is no longer unseen; it has become part of your training procedure. Do this enough times, and you can overfit it just as you might have overfit the training set, and you have lost your proxy for unseen data. This is true of any held-out set you reuse this way, including a validation set, which is reused by design. A puzzle at the heart of machine learning Real machine learning research looks exactly like the iterative improvement loop we just described. Everyone gauges performance using a handful of benchmark datasets that go unrevised for years. The research community repeats an enormous, distributed loop: evaluate a model on the benchmark, revise the training procedure, re-evaluate, publish, and let the next group eke out a little more improvement. This is precisely the kind of hill-climbing against a held-out set that, by the textbook account, ought to produce rampant overfitting. By now, the leader…
Discovery · Original source
DFM scientist collaboration program launches
PhAI Labs opened a scientist collaboration program focused first on life science and biomedicine, inviting collaborators with scientific questions, data, computational or experimental environments, and real validation conditions.
PhAI Labs · Original source

AI4Science Daily · 2026-09-14 — Teaching future scientists to interrogate AI tools for scientific discovery

Thu, 24 Sep 2026 10:19:29 +0000

Selected developments

Teaching future scientists to interrogate AI tools for scientific discovery
AutoDiscovery, our AI agent for scientific research, analyzes datasets, proposes hypotheses, runs experiments to test them, and ranks the results by Bayesian surprise—the gap between what the model predicted and what the data showed. In a University of Washington classroom this spring, it surfaced a possible flaw in how battery aging is measured, an unexpected pattern in silicon polymers, and a recurring feature in proteins that respond to light. But a surprising result is not necessarily an important one—or one that will withstand further testing. Students still had to scrutinize the evidence, compare the findings with published research, and decide whether each result reflected a genuine discovery, a coincidence, or a problem with the data. Luna Yue Huang, an associate teaching professor of materials science and engineering at the UW, invited students to explore this way of working with AI through the GenAI for Science: Ai2–UW Materials Challenge. Twenty-five teams proposed a dataset and research question; ten were selected to spend several weeks working with AutoDiscovery. A UW instructional team and Ai2 researchers reviewed the proposals based on the richness and scientific distinctiveness of the data, their potential to produce meaningful findings, and the overall diversity of topics represented. Ten projects were selected, with students from other proposals joining those teams. In some cases, instructors also helped narrow broad datasets to a more manageable scope aligned with the students’ research interests. Students tested AutoDiscovery’s hypotheses against the available evidence, identified gaps in its reasoning, and assessed whether it meaningfully reduced the amount of manual data analysis required. AutoDiscovery's performance varied considerably—in some cases, it surfaced leads the students might not otherwise have explored. Here are four representative projects from the class: Project 1: Lithium battery aging A group studying lithium-ion battery aging surfaced a counterintuitive lead: that the periodic diagnostic tests used to check a cell's health may themselves speed its degradation. This rested on a correlation in the data rather than proof that the tests caused the extra wear, so confirming it would take a controlled experiment—a clear and logical next step for follow-on work. Project 2: Estimating an atom’s charge Another group, drawing o…
Allen Institute for AI · AI for Science · Original source
HypoKG: Evidence-Disciplined Biomedical Hypothesis Generation Beyond Endpoint Knowledge
Large language models (LLMs) can generate biomedical hypotheses, but it remains unclear whether they truly reason from scientific evidence or simply produce convincing-sounding ideas. To study this, we combine three major biological databases: the Kyoto Encyclopedia of Genes and Genomes (KEGG), Rhea, and UniProt, into a unified biochemical knowledge graph and construct a benchmark of 550 paths connecting enzyme sources to rare disease endpoints, yielding 13,200 hypotheses from six LLMs under four conditions varying the biological information each model receives: source enzyme only, full biological path, or source and disease endpoint only. Hypotheses are scored using an expert-derived five-criterion rubric on a 1-5 scale per criterion. We find that models given both the source and disease endpoint often produce the highest-scoring hypotheses, showing that LLMs can generate compelling ideas from minimal information. However, these hypotheses are less grounded in the evidence. In contrast, models given the full biological path generate hypotheses more consistent with known mechanistic relationships. We call this evidence-disciplined reasoning. To confirm this effect, we shuffled intermediate path steps while keeping endpoints fixed. Evidence grounding dropped significantly (delta = -0.793, p < 0.001), confirming models genuinely used path structure during reasoning. Our findings show that knowledge graphs support hypothesis generation in two ways: they identify biological endpoint pairs absent from the literature, and their mechanistic paths guide how LLMs reason between them.
arXiv Quantitative Methods · Original source
'Multifunctional' brain implant translates speech and gestures in real time
Human communication relies on more than just talking; we nod, shrug, move our hands and point with our fingers. For people with severe paralysis, brain–computer interfaces (BCIs) can convey intended speech or movement. Now, researchers have shown that a single brain implant can do both at once. The device, described1 today in Nature Neuroscience, is the first to translate both verbal and non-verbal modes of communication, simultaneously and within seconds of the user’s intent. It uses artificial intelligence to convert electrical brain activity into text that appears on screen and to prompt a personalized animated avatar to move. Recent BCI technology has enabled people with paralysis to do tasks such as talking2 and controlling a cursor3 or a robotic arm4. “But most of these studies have focused on restoring one of these functions at a time,” says study co-author Samantha Brosler, a PhD student in bioengineering at the University of California, San Francisco. This system is a proof of concept but aims to be “a step towards more multifunctional BCIs that can capture more of the expressivity and communication that is possible” than with current technologies, she adds. “This gives us hope for neural prosthesis that will enter day-to-day usage,” says Christian Herff, a computational neuroscientist at Maastricht University in the Netherlands. Multitasking challenges The difficulty of decoding gestures and speech at the same time is that some of the neural signals that are involved overlap. And the patterns of neural activity can be slightly different during tasks that involve simultaneous speech and gesturing, compared with that during speech-only or gesture-only tasks. To distinguish between the two, Brosler and her colleagues worked with two participants. The device — a surgically implanted array of 253 electrodes placed on the surface of the brain’s cortex — “covers a fairly large surface area of the sensorimotor cortex, so it can capture signals related to both speech and body movements”, explains Brosler. The first participant had impaired speech and movement following a brainstem stroke. The researchers asked him to silently speak five phrases and attempt to wave, nod, shake his hands and clap. The second participant had a type of motor neuron disease called amyotrophic lateral sclerosis, which caused progressive speech paralysis and motor impairments. At…
Nature Portfolio · Nature · Original source
Pick Your Poison: Learning to Select Poison Sets for Stronger LLM Backdoor Attacks
Backdoor poisoning attacks add poisoned examples to otherwise-clean finetuning data, pairing a trigger with a target behavior that the model learns to produce when the trigger appears. Existing evaluations typically fix the number of poisoned examples and sample them at random from a candidate pool. We show that this can severely underestimate worst-case vulnerability: across three LLaMA-3-8B backdoor settings, holding the model, clean data, and poison count fixed, attack success ranges from 3% to 80% depending only on which poison set is chosen. We formalize poison selection as oracle-budgeted set optimization and introduce SAILS (Set-level Audit-Informed Iterative Learned Selection), which learns a set scorer from a few hundred finetune-and-evaluate runs, ranks millions of candidate sets, and audits only a small shortlist. SAILS improves held-out attack success by 30 percentage points on average over the strongest influence baselines, transfers from small-scale to full-scale finetuning, and extends to code-generation, agentic, and API-only backdoors.
Hugging Face Daily Papers · Original source
RSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments
Digital agents must often adapt to new environments whose interfaces, tools, and failure modes are not fully captured by pretrained models. We introduce RSIAgent, a training-free multi-agent framework for recursive self-improvement through autonomous memory construction. RSIAgent coordinates curriculum, actor, and verifier agents to continually explore the environment, validate outcomes, and retain environment-specific knowledge, including reusable causal relationships between actions, conditions, and consequences. It further adopts a broad-then-deep exploration strategy, combining parallel broad recursive self-exploration for discovering diverse environment structures with focused deep self-exploration for uncovering hard cases, hidden constraints, boundary conditions, and previously unknown causal dependencies. The resulting memory is frozen and can be directly reused for downstream tasks without updating model parameters. Experiments on OSWorld-v2 and Agent's Last Exam show that RSIAgent substantially improves strong open-source models, enabling Kimi-K3 and GLM-5.3 to outperform frontier closed-source models including GPT-6.
Hugging Face Daily Papers · Original source

AI4Science Daily · 2026-09-13 — 25 Fields Medalists, Including Terence Tao and Deng Yu, Warn of "Severe Misalignment" Between AI Firms and Mathematics

Sat, 10 Oct 2026 10:19:22 +0000

Selected developments

25 Fields Medalists, Including Terence Tao and Deng Yu, Warn of "Severe Misalignment" Between AI Firms and Mathematics
Twenty-five Fields Medalists, with Terence Tao and Deng Yu among the first signers, issued an urgent public statement arguing that the way AI companies are pursuing mathematics is now in "severe misalignment" with the goals of the mathematical community. The statement explicitly does not call for stopping AI or deny its achievements. It raises three concerns: AI is solving problems faster than the community can digest the knowledge; AI firms optimize for benchmarks while mathematics values methods, understanding and transmission of knowledge; and once open problems are solved quickly, priority, training and open culture need new rules. Triggering the letter was OpenAI's claim of a Navier-Stokes result produced by 10,000 agents running for 88 hours.
量子位 · Original source
Spontaneous thought transitions follow a foraging-like policy on high-order memory patches
When people engage in spontaneous thought under low external constraint, thought content changes continuously and can shift across topics. However, the cognitive computational principles governing these transitions remain unclear, as does whether similar dynamics emerge in artificial language generation. Here, we formalized spontaneous narrative dynamics using a foraging-inspired framework, constructing an agent based on the marginal value theorem (MVT) that balances exploration and exploitation while selecting among informational resources represented in a participant-specific, high-order patch network. Patches were defined as partially overlapping composite structures containing second-order paths and third-order interactions, with their available resources quantified using multilevel structure values. In a multi-theme free spoken narrative task, the model predicted the temporal organization of human patch activation substantially better than random-shuffle baselines, although the strength of model conformity varied across participant-theme samples. Model-predicted patch transitions corresponded to systematic semantic segmentation and reorganization across multiple representational levels. Transition rates were positively associated with the perceived importance and likelihood of the narrated events. These findings were replicated when the modeled memory environment was enriched with additional hub-related narrative resources in a supplemental experiment. Control-model analyses further showed that patch-network topology, resource depletion history, or either order of structure value alone could explain only part of the observed dynamics; the most stable human prediction required foraging-related decisions guided jointly by multilevel structure values. Narratives generated by large language models (LLMs) also showed detectable conformity with the same framework, but the full model captured LLM content less consistently, while simpler control models accounted for a larger proportion of their patch-level dynamics. Together, these findings support a computational account in which patch-level narrative dynamics are organized by value-guided foraging over structured memory resources, while revealing graded differences in the strength and stability of these dynamics across human and artificial language generation.
bioRxiv Latest · Original source
E2A-Bench: Benchmarking Evidence-to-Action Reliability in Financial Chart Reasoning
Can financial vision-language models (VLMs) turn chart evidence into reliable action recommendations? Existing hallucination evaluations are mostly claim-centric; they assess whether generated statements are supported, but not whether evidence remains traceable through rationale, confidence, and final action. We introduce E2A-Bench, a 969-query benchmark for financial chart reasoning, constructed from 323 HS300 constituents under three input modalities with deterministic OHLCV-derived evidence anchors. E2A-Bench evaluates grounding, reasoning-action consistency, evidence-confidence calibration, and directional coverage through UCR, RCI, ECI, and NDR, where NDR measures coverage-aware evidence-to-action reliability rather than realized trading performance. Evaluating 20 VLMs reveals three failures hidden by scalar hallucination scores: the lowest-UCR model ranks near the bottom by NDR due to only 6.4% directional coverage; oracle-aided verification reduces unsupported claims but can collapse coverage; and financial fine-tuning amplifies the BUY:SELL ratio by factors of 4.21 to 4.68 across strict base-fine-tuned pairs. These results show that financial VLM evaluation should trace the full evidence-to-action chain rather than rely on a single hallucination score. Code and data: https://github.com/wanng-ide/E2A-Bench
Hugging Face Daily Papers · Original source
Continuous Diagnostic Indices from BanffNET Automated Lesion Scores for Kidney Transplant Pathology: Development and Evaluation of a Diagnostic Prediction Model
Background: Histopathological diagnoses in kidney transplant biopsies, as described by the Banff classification, suffer from substantial inter-observer variability and misinterpretation, making them prone to diagnostic errors. Continuous indices based on BanffNET automated lesion scores are proposed to quantify the continuous phenotypic spectrum of kidney transplant biopsies, complementary to pathologist assessment of biopsies. Methods: The BanffNET model was used to obtain continuous lesion scoring for a training cohort of 2544 biopsies, a large validation cohort of 3863 biopsies, and a multi-reader cohort of 36 biopsies scored by 67 pathologists. A penalized regression approach was used to condense 17 automated lesion scores into four indices, representing the most important dimensions of graft injury: a TCMR/TI Index, an AMR/MVI Index, an Activity Index and a Chronicity Index. The BanffNET Indices were evaluated against Banff diagnoses and histological indices in the training cohort and validation cohort. Since pathologist-assigned Banff lesions and diagnoses suffer from inter-observer variability, associations were also assessed with diagnoses in a multi-reader study, with time to kidney graft failure in the training and validation cohort and with biopsy-based molecular signatures in the validation cohort. Results: The BanffNET TCMR/TI Index, AMR/MVI Index and Activity Index showed excellent discrimination of Banff TCMR, Banff MVI and Banff Any Diagnosis respectively, with validation AUCs ranging from 0.80 to 0.92. In addition, the BanffNET Indices explained large amounts of variability in multi-observer reference standards, time to graft failure and molecular biopsy-based diagnostics. In the large majority of cases, variability explained by BanffNET Indices matched or exceeded the variability explained by pathologist-assigned diagnoses and histological indices. Conclusions: BanffNET Indices are reproducible descriptors of whole slide images that could complement and augment kidney transplant biopsy evaluation by pathologists. Associations of BanffNET Indices with multi-observer diagnoses and related outcomes suggest that they capture additional information on tissue injury type and severity on top of pathologist-assigned Banff diagnoses and histological indices.
medRxiv Latest · Original source

AI4Science Daily · 2026-09-12 — GPT-6 Astra Clears the Last FrontierMath Tier 4 Problem, Prompting Epoch AI to Call the Tier Saturated

Sat, 10 Oct 2026 10:19:22 +0000

Selected developments

GPT-6 Astra Clears the Last FrontierMath Tier 4 Problem, Prompting Epoch AI to Call the Tier Saturated
Epoch AI says FrontierMath Tier 4 is now "saturated": by its accounting, when attempts across different models and different time periods are accumulated, every Tier 4 problem has been correctly solved by an AI at least once. OpenAI reports a score of 97.6% for GPT-6 Astra, which solved the one problem that had never been cracked before. When Tier 4 launched on July 11, 2025, the top score on the board was only about 5%. Jay Pantone, an associate professor of mathematics at Marquette University and one of the problem setters, said earlier AIs tended to find numerical shortcuts, whereas this solution was quite close to his own. Tier 4 was later revised: a v2 released in June 2026 corrected 12 problems and removed 7, leaving 43. The project has since added Open Problems and FrontierMath Erdős, where Astra solved only 2 of 68 problems.
量子位 · Original source
Thought without Systematicity? Evaluating Reasoning Models on Rule Induction Tasks
A central tenet of human cognition is systematicity, the principle that understanding one concept is inherently tied to understanding close variations of that concept. Do reasoning models robustly exhibit such systematicity? If so, we would expect consistent performance on structurally equivalent variants of the same task. Here, we extend established rule induction tasks from cognitive science to assess the systematicity of thought in current reasoning models. Each task family has compositional structure that we use to create structurally equivalent task variations through task isomorphisms such as recombination and substitution. We find that despite being able to correctly solve a task, models often fail on structurally equivalent variants of the same task. These findings suggest that many model behaviors lack systematicity, rendering it difficult to robustly establish the cognitive abilities of reasoning models beyond the particular contexts they were evaluated in.
Hugging Face Daily Papers · Original source
SAIR competition: Andrews–Curtis challenge
[This is a guest post by Lucas Fagan . This blog post was initially written in a different file format and converted using AI. — T.] I am excited to announce the Andrews–Curtis Conjecture Challenge , which opens today. This challenge is a collaboration between the SAIR Foundation and the Math-AI group at Caltech , organized by Sergei Gukov, Terence Tao, and myself. The Andrews–Curtis conjecture is one of the most prominent open problems in combinatorial group theory and also has deep connections to low-dimensional topology. Its potential counterexamples are relevant to the search for exotic smooth four-spheres and the smooth four-dimensional Poincaré conjecture , as well as the Generalized Property R conjecture about surgery on links . Yet unlike many open problems at its level, Andrews–Curtis can be formulated as a combinatorial search problem with easily checkable solutions, making it ideal for a challenge of this form. At Caltech, we have been developing reinforcement learning and combinatorial search methods for this problem to resolve potential counterexamples (see What makes math problems hard for reinforcement learning: a case study and The Two-Hump Problem ). However, many important cases have resisted all our efforts. We hope that this challenge will lead to resolving these and more (or disproving the conjecture), especially given the recent progress in AI. We give more details about the different tracks of the competition below. To briefly introduce the problem: the Andrews–Curtis conjecture says that any balanced presentation of the trivial group can be transformed to the trivial presentation using the following moves (which do not change the underlying group): (AC1) Invert a relator: . (AC2) Multiply a relator by another: for some . (AC3) Conjugate a relator by a generator or its inverse: for some . Two presentations connected by these moves are called AC-equivalent ; a presentation that is AC-equivalent to the trivial presentation is called AC-trivial . As a simple example, is AC-trivial: It is generally suspected that the conjecture is false ; there are many simple potential counterexamples in which all computational efforts have failed to find a path to the trivial presentation. The most notable of these is the Akbulut–Kirby family whose AC-triviality is open for . Indeed, , with total relator length , is the shortest possible counterexample…
Terence Tao · Original source
‘Immature playground boasting’: Mathematicians uneasy at OpenAI’s latest scalp
It was a week that left mathematicians reeling. Hot on the heels of a flurry of cases of artificial intelligence furthering the field, OpenAI declared a major scalp: its latest AI model had cracked a Millennium Prize Problem, a puzzle with a $1m reward that had defied human brains for decades. The achievement bore little resemblance to how mathematical problems normally fall. A near-trillion dollar private company had unleashed 10,000 agents – AI systems that carry out tasks autonomously – on the problem. The bill was estimated at $15m. The trajectory towards AI capable of superhuman maths has long been clear, but to nail such a substantial problem so swiftly sent shock waves through the field. Mathematicians are asking what will be left for them if works in progress are hoovered up and claimed by others, and how they should train the next generation when even fiendish assignments can be solved at the press of a button. “I feel slightly shell-shocked,” said Prof Colva Roney-Dougal, the head of pure mathematics at the University of St Andrews. At a recent public lecture, she said she believed that AI was unlikely to do anything remarkable soon. “Three months later, I’m totally wrong,” she says. “We’re all just waiting to see what happens. It’s coming so fast.” Prof David Silvester, a mathematician at the University of Manchester, said the field felt “very unstable” given the pace of change. “All the open maths problems could fall with enough resources,” he said. “This is irreversible. This is not going to change.” Prof James Robinson, a mathematician at the University of Warwick, said hard problems were the fuel of mathematics, driving creative approaches across successive generations. “It’s frustrating to see these big tech companies burning this fuel up just so that they can show off about how great their latest model is,” he said. “It seems like immature playground boasting writ large, underpinned by billions of dollars and the potential for significant environmental damage in an age when climate change is probably the biggest challenge we face.” One reason Silvester went into maths was the rush that comes from cracking a problem. For many mathematicians it is the weeks, months and years spent circling a problem, breaking it down, trying one approach after another, and never quitting that appeals. “Without that, the subject seems completely different to m…
Discovery · OpenAI · Original source