Cross-disciplinary collection
AI4Science field
Research releases and public discussion in this scientific field, ordered by time.
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 sett
A transformer language model performs a bounded amount of computation per token, and recent work by Vishal Sikka, former CEO of Infosys, argues that this bound limits which tasks a model can carry out or verify (arXiv:2507.07505). We ask how much of the budget beneath that ceiling is spent on work the model has already done. Serving is stateless across requests: a model that answers a second question about a document
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.
Self-driving laboratories (SDLs) combine automated experimentation with adaptive decision-making to accelerate scientific discovery. Their operation nevertheless often depends on human specialists who translate scientific objectives into executable closed-loop campaigns. Specialists adjust them as data and operating conditions change. Here, we present La Agente \'Optima, an agentic framework that constructs and super
Abstract High-quality multimodal datasets are essential for developing vision-language models, yet publicly available figure-text resources in specialized scientific domains remain limited. To address this gap, we present a large-scale figure-text pair dataset constructed from map-related scientific literature (FTPD-ML). The dataset was derived from 96,859 scientific publications in cartography, geography, remote sen
How humans grow and maintain abstract knowledge from the sparse, streaming noisy data of experience is a longstanding challenge in cognitive science. Any computational account must satisfy at least three desiderata: It must be (1) data-efficient and compute-efficient, (2) capture gradations of uncertainty to support intelligent inquiry and information gathering, and (3) be flexible enough to mentally represent the en
Computer science papers rely heavily on diagrams: architecture drawings, system flowcharts, and pipeline schematics that often carry more information than the text around them. There is currently no public dataset that pairs this specific kind of figure with captions, context, questions, answers, and step-by-step reasoning, which is exactly what is needed to train a vision-language model to understand them. We presen
AI can already assist with a surprising amount of scientific work: search literature, analyze datasets, write code, and generate and test hypotheses. The harder problem is building systems scientists can steer as a research project unfolds, especially when new evidence changes which hypotheses to pursue, experiments to run, or questions to investigate next. That challenge was the focus of an August 27 event at Ai2 ma
Over the course of the last decade, neural networks have grown from an academic curiosity to moving the markets of nations. Despite this explosion in both research and deployment, relatively little is understood about how they achieve the solutions they do. This is both scientifically relevant, and pressing for society. When neural networks make decisions across self-driving, construction, law, hiring and health, the
Nature Biotechnology, Published online: 01 September 2026; doi:10.1038/s41587-026-03267-1 Reference materials should be adopted as a common calibrator for multiomics measurement and co‑profiled with study samples. Multiomics results should be reported as sample‑to‑reference ratios so that they are reproducible and suitable for artificial intelligence tools.
Post-training quantization is often treated as a semantically neutral optimization for edge deployment of Large Language Models. When a full-precision source checkpoint is evaluated and quantization is applied downstream without equivalent re-evaluation, this workflow creates a structural validation--deployment gap: because quantization is a many-to-one mapping over parameter space, source-precision certification doe
As AI begins to build itself, automating alignment research becomes increasingly important to let safety research keep pace. Although measuring the success of alignment research is enormously challenging, researchers (at Anthropic and elsewhere) have developed benchmarks and automated auditing tools, such as Petri, that quantify common alignment failures, like deception, sycophancy, and jailbreaks. In one of our earl
PhysicsNeMo General Release v2.2.0 Added Adds ShardTensor support for GeoTransolver and FLARE models. Adds zarr save/load for Mesh and DomainMesh via tensordict's zarr storage backend: physicsnemo.mesh.io.to_zarr / from_zarr , with training-appropriate chunking and zstd compression. MeshReader and DomainMeshReader transparently read zarr stores alongside .pmsh / .pdmsh (opt in via pattern ). Requires optional zarr >=
Reinforcement learning (RL) post-training has emerged as a powerful framework for enhancing the capabilities of large language models (LLMs), enabling impressive reasoning, math, and coding capabilities. Yet for many researchers and practitioners, the principles behind classical RL remain a "black box". In this work, we deconstruct the RL post-training algorithm, investigating each step to clarify what is actually ha
When a large language model (LLM) is asked to write a person's life, how much of what it writes actually happened? We present a scene-level case-study audit - the first quantified audit of LLM-generated autobiography against a subject-specific ground-truth corpus that we are aware of, based on an unsystematic literature search. The subject and the author of this paper are the same person: a 366-day "page-a-day" book
Large language models (LLMs) are becoming routine instruments of scientific research, assisting with literature synthesis, hypothesis development, coding, and formal reasoning. Their use raises a central epistemic question: when parts of scientific reasoning are delegated to an artificial system, what conditions must remain under human control for the resulting knowledge claims to retain epistemic legitimacy and acco
The rapid evolution of energy structures has positioned microgrids as pivotal components of next-generation power systems, offering enhanced resilience and renewable energy integration. However, the inherent low inertia, complex dynamics, and poor model conditions of microgrids necessitate advanced data-driven frequency control strategies. Although reinforcement learning (RL) has demonstrated certain potential and ad
Transport resilience work is often split across separate data preparation scripts, network models, simulation tools, image inspection systems and reports. This fragmentation makes it difficult to move from an observation to a tested and reviewable decision. We introduce ResiliFlow, an open transport world model concept and an implemented platform for infrastructure resilience, response and recovery. The platform conn
Graph neural networks (GNNs) are routinely employed for short-range forecasting on multivariate time series with a spatial graph structure. Despite the availability of many alternative datasets, method innovations within this domain are predominantly assessed against a rather limited set of benchmark datasets, most notably Chickenpox, PedalMe, WikiMaths, METR-LA, and PEMS-BAY. The evaluation protocols contain baselin
Nature Machine Intelligence, Published online: 24 August 2026; doi:10.1038/s42256-026-01290-0 Implicit neural representations can turn massive scientific measurements into compact functions. A hierarchical approach now helps these representations preserve richer fine-scale detail.
Scientific data acquisition continues to outpace storage and analysis capabilities, making voxel-based representations increasingly intractable. Implicit neural representations (INRs) offer a promising solution by encoding signals through coordinate-based neural networks, serving as surrogates of data, with computational and storage requirements scaling with network complexity rather than data dimensionality. However
Researchers at the UK AI Security Institute used psychometric methods to show that popular safety benchmarks for language models don't measure one consistent trait. Blanket blocking of requests can artificially inflate a safety score even as the model gets less useful day to day. The study also offers a method for catching models that act more cautious during tests than they do in normal use.
Mechanistic interpretability aims to read neural networks like programs, but superposition forces features into low-dimensional projections. This study trains a one-layer transformer and decomposes its virtual weights between tokens, positions, features, and logits, scoring weights by effectiveness and helpfulness. They find an interference weight: the path from 'IN' to logits via residual stream votes for 'utions',
This tutorial introduces AutoFigure, a practical toolkit for generating scientific figures directly from text descriptions, paper-like content, and structured methodological explanations. Steps include setting up the AutoFigure environment, fixing dependency issues such as Pillow compatibility, preparing SVG and PNG rendering tools, building a custom reference figure, configuring an API-backed generation workflow, an
New Nvidia research highlights that for long-horizon tasks, the software harness—the tools, memory management, and rules around a model—matters more than the model itself. Using a custom harness with improved memory handling and a supervisor component, researchers achieved a 100% score on the interactive reasoning benchmark ARC-AGI-3 with Claude Opus 5, compared to 30% without the harness, which was still the best am
Google Research researchers introduce Mobility-Embedded POIs (ME-POIs), a dynamic, mobility-informed framework that enhances language models' understanding of places by integrating aggregate, anonymized mobility patterns such as arrival times, stay durations, and surrounding movement. This self-supervised approach blends text descriptions with large-scale mobility data to create embeddings that capture both a place's
Researchers at Google Research introduce the Biomarker Discovery Framework, a multi-agent system that prioritizes candidate biomarkers from wearable sensor data through iterative hypothesis generation, statistical analysis, adversarial validation, and literature-grounded reasoning, under human supervision. The system combines deterministic computation for numerical analysis with generative reasoning for hypothesis fo
This survey proposes viewing agent evolution as dynamic graph transformation, modeling agent state as a dynamic graph where memories, tools, skills, workflows, and inter-agent relations are represented as typed nodes, edges, and subgraphs updated through schema-constrained rewrites. The authors organize existing dynamic-graph-based methods for self-evolving agents into four taxonomies: node/feature evolution, edge/to
Robustness verification of neural networks is increasingly important, due to their use in many critical domains. In certain scenarios, proof sharing has been shown to accelerate incomplete verification techniques by reusing intermediate-layer abstract states, or templates, across queries. However, questions remain as to the robustness of template-based acceleration across varying network architectures, properties, da
Humyn Labs has unveiled an egocentric video library for physical AI, comprising 5 first-person datasets categorized by environment and capture type. The datasets feature 6-DoF head poses, 21-point hand keypoints, absolute metric depth, and per-frame action labels. Data is delivered in MCAP, RLDS, and LeRobot v3 formats.
Ultra-FineWeb-L1, part of the UltraData ecosystem, is now open-sourced, offering over 1 trillion tokens of high-quality English web data processed from Common Crawl. The release underscores the importance of high-quality data for training powerful large language models (LLMs).
This tutorial presents an end-to-end preference-learning workflow using the Anthropic HH-RLHF dataset and Direct Preference Optimization (DPO). It covers preparing a Colab environment, loading and parsing chosen-rejected response pairs, auditing the dataset for structural and length-based biases, and performing lexical shortcut diagnostics. The process includes tokenizer-aware length filtering and a version-robust DP
EXIMO is an efficient algorithm for fine-tuning vision-language-action (VLA) policies to learn new tasks quickly. It operates in three stages: explore, imitate, and optimize. During exploration, a vision-language model (VLM) acts as a planner, breaking down long-horizon tasks into shorter subtasks for the VLA, and together they collect an orchestrated dataset. During imitation, the VLA is fine-tuned on this data. Fin
Large language models often fail to answer questions about a bounded document collection without retrieval at inference. This work frames the problem as document knowledge internalization, converting a fixed corpus into usable parametric knowledge. The proposed IAR (Inject, Align, Recover) framework separates structured knowledge injection, QA alignment, and general ability recovery. Inject uses continuation, rewrite
Nature Computational Science, Published online: 20 August 2026; doi:10.1038/s43588-026-01035-4 As large language models begin to plan and run experiments on their own, the reflex is to demand that they be interpretable before they are trusted. We argue that what grounds trust in autonomous science is not a view inside the model; it is provenance: a complete, re-openable record of what was reasoned, done and measured
Customer-service LLM agents must follow organizational policy when acting on a user's behalf. Compliance failures arise from either forbidden actions, such as granting an ineligible change, or omitted procedural requirements, such as identification or confirmation. Runtime safeguards can intervene on risky actions, but action-local checks do not guide an agent through a multi-step procedure. PolicyGuide compiles each
Public benchmarks are important for measuring Automatic Speech Recognition (ASR) capabilities, but models may be optimized for these benchmarks in ways that fail to generalize. This paper presents a methodology to quantify such benchmark optimization, focusing on cases where audio underdetermines the reference transcript. The authors identify three families of behavioral probes—reference disagreement, masked-number r
Self-supervised learning has driven audio representation learning, but existing methods often rely on complex pre-training recipes. The authors introduce NAPE, a framework where a causal Transformer predicts the next patch embedding of a log-mel spectrogram from previous ones, using only causal masking and stop-gradient as training signals, avoiding reconstruction decoders, acoustic tokenizers, and other components.
Existing code generation systems often assume a predefined repository architecture, which fails in zero-to-all code generation where an agent must build a complete software project from natural-language requirements while maintaining modular architecture. We present Repo0, a continuous structural evolution framework that maintains an explicit architectural state as a Dual-Directed-Acyclic-Graph (Dual-DAG), combining
LLM agents learn by interacting with environments, yet existing environments are hand-built and static: blind to an agent's weaknesses, and quickly left behind as it improves. While recent environment generation methods attempt to address this, they require domain-specific pipelines, rely on expensive or unreliable verifiers, and still produce static environments. To alleviate the engineering burden, we propose Envir
MemTrapBench is a new benchmark for evaluating cognitive traps in large language model (LLM) memory use. Existing memory benchmarks mainly assess whether information is correctly extracted, stored, and retrieved, but overlook how retrieved memories reshape model reasoning and affect current task performance. MemTrapBench covers two forms of cognitive traps: Reasoning Fixation and Belief Distortion. Experiments across
Cross-lingual knowledge transfer is critical for multilingual models with scarce target-language data. Existing methods require parallel data, translation systems, or extra training stages. The paper proposes LINK, a data-level intervention that substitutes words in high-resource (English) pretraining data with translations using bilingual vocabularies, requiring no extra training. Evaluations on eight languages and
A study published in PNAS found that X's recommendation algorithm tends to prioritize content inconsistent with users' values, i.e., anger-inducing content, to increase engagement, forming a cycle of 'more anger, more interaction, more recommendations.' Conducted by Stanford and other institutions, the study involved 715 U.S. users who installed browser extensions for data collection and completed value surveys. The
GEqTrain is a configuration-driven framework that separates dataset semantics, model composition, and training objectives, enabling reuse of equivariant graph neural networks across diverse 3D scientific tasks. Raw data are mapped to typed fields, and models, losses, and workflows are assembled declaratively via Hydra configurations. The framework is demonstrated on three tasks within one software stack: coarse-grain
SPADE is a self-play RL framework where a single LLM acts as both an Environment Designer, which writes complete, long-horizon training environments as executable code with an OpenAI Gym-style reset()/step() interface, and a Reasoning Agent that learns to act in them. The designer optimizes the agent's regret signal (the gap between reward with and without privileged hints) to generate environments at the edge of cap
Reward modeling for joint video-audio generation typically combines metrics for audio quality, visual fidelity, and synchronization, which evaluate perceptual dimensions separately and fail to capture the overall semantic and temporal coherence among text, video, and audio, encouraging reward hacking. To address this, the authors construct a large-scale human-preference dataset VAPref-10K with 9K prompts and 10.3K fi
Single-step retrosynthesis is central to computer-aided synthesis planning, but its one-to-many nature is poorly captured by single-answer benchmarks. To address this, the authors introduce Top-K prompting as a training and inference paradigm for diverse, plausible predictions. They compile CREED-CCV-2+USPTO-XL, an ultra-large-scale dataset of ~45.6 million verified reactions, to train C3LM (Chemistry Constraint-Cons
This paper presents a multi-dimensional analysis of human-like behaviors in large language models (LLMs), focusing on their prevalence, potential effects, and controllability. Using LLM-as-a-judge and human evaluation, the authors analyzed 21,000 multi-turn conversations from four widely used models: gpt-4o, gpt-4.1-mini, claude-sonnet-4.6, and gemini-2.5-flash. They found that human-like behaviors are pervasive but
We introduce PTXBench, a benchmark for evaluating and adapting large language models (LLMs) to use architecture-specific PTX for GPU kernel optimization. PTXBench measures functional correctness, whether selected target instructions execute at runtime, and speedup over frontier libraries across GEMM and attention workloads on H100 and B200 GPUs. Our evaluation shows that architecture-specific PTX capability remains u
LEGO-RL is a framework that bridges native coding-agent harnesses with scalable policy-gradient optimization without modifying their internal control flow. It ensures faithful optimization via in-process LLM proxying for token-level alignment, reliable execution through scalable sandbox orchestration with image caching and stage-wise defenses against reward hacking, and observable training via an integrated plugin an
We introduce LifeSciBench, a benchmark of 750 expert-authored tasks designed to evaluate whether language models can handle realistic life science research work. The majority of existing life sciences benchmarks have a narrow scope or are purely knowledge-based, and therefore fail to capture the complexity of real-world research, which often involves ambiguities and requires the accurate execution of multiple depende
We introduce Semantic Task Completion Video Generation, an outcome-oriented task requiring both the achievement of the intended outcome and semantic grounding, which characterizes the correspondence between the reference image and the generated outcome in terms of high-level task-relevant semantics. Evaluation focuses on the generated outcome and does not require a complete sequence of intermediate steps or appearanc
Modern agents operate within agent harnesses that manage tools, context, and control flow. The original Agent Lightning proposed a disaggregated architecture connecting arbitrary agents to RL training via an LLM endpoint proxy—a paradigm termed harnessed agentic RL. v1.0 is a lightweight framework of about 3,500 lines of code, supporting arbitrary harnesses and offering a reproducible pipeline for coding-agent RL. Wi
A study published online in Scientific Data on August 18, 2026, presents a dataset of red blood cells captured using digital holography, designed for machine and deep learning analysis of hereditary anemias. This resource provides a standardized collection to support the development and application of algorithms for identifying and classifying morphological features associated with these conditions, facilitating comp
Google Research demonstrates the feasibility of PhotoScan, a deep learning approach that estimates body composition from smartphone photos to predict insulin resistance with accuracy comparable to DXA scans in a clinical research setting. Insulin resistance, a critical driver of metabolic disease, often precedes type 2 diabetes by years, and a HOMA-IR score above 2.9 indicates insulin resistance. The study highlights
Florent Delgrange won the Best Blue Sky Paper Award at AAMAS 2026 for his work 'Foundation World Models for Agents that Learn, Verify, and Adapt Reliably Beyond Static Environments.' The paper proposes a research agenda that merges reinforcement learning and reactive synthesis into a single loop: a foundation world model is learned from experience but structured for a verifier to reason about. As the agent learns a p
Neurodegenexrative diseases such as Alzheimer's disease and Parkinson's disease are diagnosed most reliably only after substantial, often irreversible, neuronal loss has already occurred, creating an urgent need for quantitative tools that can detect subtle, early, and individual-specific brain changes from neuroimaging data. This review surveys a broad and rapidly evolving toolkit of data-driven techniques for trans
Post-training papers, model cards, and blog posts often treat scores on a small set of coding benchmarks (e.g., SWE-bench and LiveCodeBench) as evidence of broad coding capability, both for research artifacts and user-facing systems. We argue that optimization for these benchmarks leads to measuring task-specific performance, creating a meaning gap between measured scores and claims of general coding ability. We exam
This paper introduces a common layer equation that unifies many graph neural network architectures via seven components: update domain, channel set, propagation bank, per-channel message maps, channel-fusion operator, ego/residual map, and update map. The central factorization separates where information moves (propagation bank) from what moves (message maps). Function-valued fillings extend the equation to local mes
Methods for improving knowledge use in LLMs typically fall into non-parametric retrieval or parametric adaptation. Engram-style hashed memory stores learned information in an external, addressable table, consumed through a small learned reader. This study investigates what matters more when such memory is transferred across backbones: the frozen memory or the target-side reader, via cross-model frozen-memory extracti
PixRestore is a VAE-free pixel-space Diffusion Transformer for unified image restoration (UIR), performing flow matching directly on patchified pixels to preserve fine details. It adapts to degradations by predicting layer feature reliability via LQ-HQ DINO similarity, fusing reliable layers as dense conditioning while supervising less reliable layers to remove degradation. Trained on a large-scale corpus and finetun
Reliable uncertainty quantification (UQ) is essential for deploying large language model (LLM) agents in complex interactive environments. Existing UQ methods often rely on local signals, such as token probabilities or predictive entropy, overlooking long-range dependencies where errors accumulate across an execution trajectory. We propose RUPA, a trajectory-level UQ framework that represents an execution history as
Resource rationality in cognitive science asks how agents should allocate limited computation to maximize expected value. Most reasoning benchmarks use independent per-task budgets, and existing shared-budget studies do not calibrate against single-problem competence. We introduce R^3-Bench, evaluating six-problem suites under shared budgets across mathematics, competitive programming, and abstract reasoning, in both
As text-to-image generative models advance, they raise safety concerns, particularly generating Not-Safe-For-Work (NSFW) content, further exacerbated by red-teaming attacks. Existing defenses often rely on white-box assumptions, while black-box alternatives based on LLM prompt rewriting fail in what we identify as the benign adversarial problem: prompts that are linguistically safe but still trigger harmful generatio
A Nature editorial argues that scientific breakthroughs are often sparked by innovative instruments, and AI's ability to help design them could be its most significant scientific contribution. Drawing on historical examples, the piece suggests that tool innovation is key to progress, and AI-assisted tool design may drive future discoveries.
Looped language models have shown promising results on reasoning benchmarks, yet their potential for agentic tool use remains largely unexplored. We study this question in compositional tool-calling settings, where models must coordinate multiple API calls, maintain intermediate state, and preserve dependencies across tool interactions. We evaluate native and retrofitted looped language models on API-Bank, BFCL, and
Byte-level hierarchical language models (LMs) have recently emerged as a robust alternative to subword tokenization. However, generating one byte at a time remains a bottleneck for inference speed. To address this, we introduce multi-byte prediction (MBP), which generates multiple bytes in parallel, speeding up inference with minimal performance impact and no additional parameters. MBP builds on the popular multi-tok
Language-specific competency (LSC) refers to the phenomenon where language models perform differently depending on the prompt language, attributed to internal misalignment of semantic representations. Existing solutions like English routing or language-balanced data have drawbacks. This paper introduces HOTFIXR, a data generation framework that probes a student model's multilingual weaknesses and generates targeted s
Multimodal large language models increasingly use visual chain-of-thought (Visual CoT), which generates intermediate reasoning images for spatial, temporal, and embodied reasoning, but introduces significant inference overhead. This paper introduces Internalized Visual Thinking (IVT), a post-training framework that jointly optimizes textual prediction and next-embedding prediction on unlabeled videos, enabling the mo
Memory is becoming core infrastructure for long-horizon LLM agents, yet existing evaluations offer limited guidance on which memory substrate—the underlying medium in which memory is represented and stored—should be used under different operating regimes. We present a controlled harness evaluation of memory substrates for memory-augmented agents, covering dense and sparse indices, text records, structural stores, hie
Audio-Text Foundation Models (ATMs) are prone to catastrophic failure under severe acoustic noise, but existing adaptation methods either reinforce noise instead of signal or require privileged noise annotations. PRISM (Prototype-Rectified Iterative Self-Supervised Manifold Denoising) is a training-free, source-free TTA framework that leverages the Affine Noise Hypothesis, positing that severe acoustic noise induces
A study published online on August 15, 2026, in Scientific Data introduces a Chinese corpus of human olfactory descriptions and entity annotations, driven by large language models (LLMs). The corpus aims to collect and annotate Chinese expressions related to smell, providing a data resource for olfactory research. The paper's DOI is 10.1038/s41597-026-08117-y.
The rapid advancement of artificial intelligence has substantially increased the demand for large-scale datasets; however, available datasets remain insufficient for data-intensive tasks, particularly in nuclear medicine. While large-scale datasets exist for computed tomography (CT), few provide paired projection-domain and image-domain data across diverse lesion types. This issue is more pronounced for positron emis
MOSS-VL is an open vision-language model family that treats real-time interaction—perceiving while speaking—as a first-class capability. It is co-designed across the stack: the language decoder attends to vision only through gated cross-attention, enabling natural perception of incoming frames during generation; a synthesized interaction corpus supervises when to speak, stay silent, or revise; and a staged curriculum
RobustiPy is a frontier open-source library for asking whether empirical scientific conclusions survive all possible yet defensible modeling choices. It builds analytical multiverses and summarizes uncertainty. Through controlled simulations, time profiling, and empirical replications—including re-analyses of discrepancies in published work—it makes robustness analysis transparent, reproducible, and interpretable.
Open-weight language models are often fine-tuned, quantized, pruned, or merged without documented provenance. We propose a data-free, white-box method for lineage verification: by removing the identity-aligned component that residual training imparts to branch products, we isolate checkpoint-specific structure across residual blocks and compute a symmetric lineage score. On residual-MLP and GPT-2 benchmarks, the scor
High-quality creative writing data for LLMs is dominated by story-centric data, limiting models' adherence to diverse creative formats. The authors propose an attribute-guided genre expansion framework that uses human-authored story prompts as creative seeds and manually curated genre attributes to enforce structural, stylistic, and formatting conventions. This produces genre-faithful query-response pairs, which are
Popular facts are memorized more deeply during pretraining and resist removal longer than rare ones, yet existing LLM unlearning methods apply uniform gradient pressure regardless of training-data frequency. We propose AdaPop (Adaptive Popularity), which combines local token confidence with a per-fact popularity-dependent exponent derived from an external proxy (e.g., Wikidata sitelinks, LLM-as-Judge), and automates
CardioState-JEPA is a cardiac foundation model that learns a shared representation across ECG, PPG, and PCG using a physiology-aware joint-embedding predictive architecture. It maps heterogeneous waveforms into a common token space, processes them with a shared Transformer encoder, and learns by predicting masked latent cardiac states. A learned delay aligner handles temporal offsets between electrical, mechanical, a
V-RAE is a video representation autoencoder that builds compact generative latents on top of frozen vision foundation model representations. A lightweight temporal pooling module removes temporal redundancy while preserving semantic structure, and a video decoder reconstructs continuous motion from compressed features. On K600, V-RAE achieves an rFVD of 2.13, outperforming all evaluated large-scale pretrained video V
Creative agents still lack an effective way to learn from high-quality human films, limiting their ability to produce cinematic-grade videos. To address this, researchers propose the Agentic Video Auto-Encoder (AVA-Encoder), a framework for learning agent-native video representations via agentic auto-encoding. AVA-Encoder transforms a video into a knowledge graph (KG) representation and reconstructs it back into vide
Nature Machine Intelligence, Published online: 12 August 2026; doi:10.1038/s42256-026-01292-y Wang et al. design efficient robot–environment interaction representations that achieve generalization across articulated robotic hand models and task adaptability in diverse cluttered grasping scenarios, suggesting a path towards general-purpose robotics.
Multiple-choice benchmarks are widely used to evaluate large language models, but MCQ scores conflate knowledge with sensitivity to option order, making them unreliable measures of model knowledge. This paper tests whether preventing the model from seeing option labels while committing to an answer removes positional influence and improves performance. Two strategies are evaluated: generation-then-matching and scorin
In this crosspost from AI Matters, a publication of ACM SIGAI, Ella Scallan interviews Akari Asai, winner of the 2025 AAAI Doctoral Dissertation Award. Asai's doctoral research focused on retrieval-augmented language models (LMs), arguing for moving beyond monolithic LMs to augmented LMs that are designed, trained, and deployed with complementary modules. She pioneered retrieval-augmented LMs, which retrieve relevant
Nature Computational Science, Published online: 10 August 2026; doi:10.1038/s43588-026-01018-5 A newly generated high-throughput experimentation (HTE) dataset is integrated with open-source data to produce a large and high-quality Buchwald–Hartwig reaction dataset for training machine learning models. The combination of HTE and active learning enabled the development of a robust machine-learning model with reliable o
This work introduces Hierarchical Self-Improvement (HSI), a framework where a single frozen LLM operates across three hierarchical scopes: a task harness, an evolver that rewrites the harness, and a meta-evolver that rewrites the evolver's strategy. The harness is task-specific and continuously evolvable via a fixed injection seam, using environment feedback. A thinking-on/off design isolates the contribution of harn
Ai2 expands collaboration with Hugging Face to accelerate open science August 6, 2026 Ai2 Ai2 and Hugging Face are expanding their partnership to make fully open AI easier to access and deploy. Hugging Face serves as a central platform for the AI community, and the expanded collaboration gives us substantially more storage and bandwidth to support our growing portfolio of models, datasets, benchmarks, and application
CogTwin, presented at IJCAI 2025, is a hybrid cognitive architecture framework designed to bring autonomous reasoning and real-time adaptation to digital twins. It addresses the gap in current digital twins that are largely reactive, relying on pre-programmed rules and static models, and thus struggle with unforeseen events. By integrating symbolic AI, sub-symbolic AI, and neuro-symbolic AI, CogTwin aims to equip dig
Microsoft introduces Echoverse, a suite of twelve training worlds for computer-use agents: ten deep domain worlds and two capability worlds, each drilling a single control (e.g., date pickers, nested filters) in varied forms. Depth is key: these worlds replicate real application behavior, come seeded with realistic data, and maintain coherent state across screens and users. Training on all twelve nearly doubles a 9B
Microsoft Research introduces EvoLib, a framework that enables large language models to learn from their own experience during inference without ground-truth labels or external feedback. EvoLib transforms past attempts into reusable skills and reflective insights, continually refining, consolidating, and reweighting them to turn instance-specific observations into increasingly general knowledge over time. This evolvi
A study led by the Australian National University (ANU) Emotions and Faces Lab has shown that humans can be trained to improve their ability to spot AI-generated faces. Deepfake faces have become so realistic that they are hard to distinguish from real photos, contributing to AI-related fraud. The training, which focuses on six perceptual qualities—distinctiveness, memorability, proportionality, symmetry, attractiven
Per-field accept/review with selective risk control at most alpha is the trust contract for document extraction, but the natural procedure silently violates it on real documents. On 13,859 genuine claude-sonnet-5 fields from 800 CORD receipts (49.0% correct), we diagnose three failure modes: document clustering (design effect 1.84–2.45), score-refit leakage (coverage 0.416 at risk 0.127, violating alpha=0.10 in 95% o
As high-quality training data becomes scarce and pretraining costs rise, Damian Borth, professor of AI at the University of St. Gallen, argues that we've overlooked a key knowledge source: already-trained models. His group's work on weight space learning treats trained neural networks as data, learning from the distilled results of millions of GPU hours of optimization rather than starting from raw data. This episode
I believe that the creation of visualization apps to illustrate mathematical or scientific concepts is a particularly favorable use case for modern coding agents, as many of the downside risks attached to other LLM use cases are limited: Not mission-critical . As such apps are not authorative sources of truth and only used for secondary purposes, a small positive error rate in the output can be acceptable. Stand-alon
Across three datasets, compliant AI analyses still produced widely dispersed conclusions sensitive to model and persona choices.
How Rust, Lean, Aeneas, and AI agents are helping scale formal verification for production cryptographic algorithms At a glance SymCrypt develops new verified cryptography using Rust, Aeneas, and Lean to provide higher security assurance. We prove that their code safely and correctly implements standard algorithms, notably for post-quantum cryptography. We are releasing verified code, specs, properties, and proofs in
Stanford GSB professor Douglas Guilbeault argues that human decision-making often relies on 'satisficing'—gathering just enough information to make a satisfactory prediction, rather than seeking optimal solutions. Experimental games show this strategy also applies to learning social rules, offering new insights into social learning and highlighting a key difference between humans and large language models (LLMs). LLM
This paper examines task construction as a control surface in continued pretraining. Standard pretraining fixes the self-supervised objective before training, with downstream feedback only available after the entire run, making the control loop coarse. The authors propose value-based pretraining (V-pretraining), which separates the learner from a lightweight task designer. The learner updates only via self-supervised
Microsoft Research AI for Science introduces Flint, a visualization intermediate language for AI-driven chart creation. Flint enables AI agents to reliably generate expressive, visually polished charts from simple, human-editable specifications. It leverages semantic data types to guide the compiler in choosing appropriate scales, baselines, formatting, and color schemes, and automatically manages layout to adapt to
Large language models (LLMs) are increasingly used as synthetic survey respondents, but their individual answers may look plausible while failing to preserve the joint distribution, latent structure, reliability, mediation pathways, and demographic effects of real human data. This study introduces a Lithuanian organizational-psychology dataset (n=263 employees; 68 items, 12 subscales) and evaluates 37 LLMs from OpenA
SkillOpt treats an agent's skill file as a trainable parameter outside a frozen target model, converting skill writing from one-shot prompting into a controlled optimization process. Across six benchmarks, seven target models, and three execution modes, SkillOpt achieves best or tied-best performance in all 52 evaluation cells, improving performance without updating model weights. The method keeps skills compact and
Token-level hallucination detectors that score each token independently often fail when the model is confidently wrong. This paper treats hallucination as a temporally extended span and frames detection as sequence labeling. Each token is scored from a 33-dimensional feature stream that fuses text statistics, Natural Language Inference (NLI) entailment, and language model surprisal, with no access to model internals.
Machines have become remarkably good at perceiving motion. Given a video, modern models can track how objects and points move through a scene with exceptionally high confidence. But perception is inherently retrospective: it explains motion that has already happened. Many of the systems and applications we want to build need tolook forward instead. A robot reaching for a cup has to anticipate how the cup will move be
HarmProfile is a content-centric benchmark dataset that collects model misbehavior across diverse harm categories and model families, defining the resulting harmful-output distribution as a model-level risk profile. It contains over 80,000 validated artifacts from 23 frontier LLMs across 13 model families, organized into 15 harm categories and 57 subcategories. Using this corpus, the authors find that frontier LLMs r
Version 3.0.3 changes the AlphaFold 3 code license from CC BY-NC-SA 4.0 to Apache 2.0. Moreover, it provides bug fixes and a performance improvement for large inputs. We thank everyone who reported issues, proposed new features, and submitted PRs! Change log Below is a selection of new features, performance improvements and bug fixes. Full change log: v3.0.2...v3.0.3 New features AlphaFold 3 is now compatible with Py
Open-source deep-learning framework for building, training, and fine-tuning deep learning models using state-of-the-art Physics-ML methods - Release v2.1.1 · NVIDIA/physicsnemo
Machine learning interatomic potentials (MLIPs) are transforming the landscape of computational chemistry and materials science. MLIPs enable atomistic simulations that combine the fidelity of computationally expensive quantum chemistry with the scaling power of AI. Yet, developers working at this intersection face a persistent challenge: a lack of robust, Pythonic toolbox for GPU-accelerated atomistic simulation. Fo
Music editing is vital in modern production, yet many systems overlook preserving musical facets that should remain unchanged. The authors define this as Music Context Preservation (MuseCP) and introduce MuseCPEval, the first evaluation framework covering four categories of music facets with fine-grained metrics. Objective validation and human studies confirm effectiveness, while case studies demonstrate utility as a
Fourteen researchers report eleven case studies across mathematics, physics, astronomy, biology and materials, documenting both concrete advances and failure modes.