Temporal Multi-Signal Fusion for Token-Level Hallucination Detection

Summary

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. A Bidirectional Gated Recurrent Unit (BiGRU) achieves an AUC of 0.840 on RAGTruth (10 seeds), an 11-point gain over an independent logistic-regression baseline (p = 0.002, Wilcoxon signed-rank). A controlled decomposition attributes most of the gain to temporal order rather than model capacity: evidence propagates from confident positions to ambiguous neighbors. The same 0.845 ceiling recurs across recurrent, state-space (Mamba), and attention architectures, identifying the feature set as the bottleneck. Since it reads only generated text and external signals, the detector works on closed-source models and maintains performance on text from unseen language models, with under 4% AUC loss.