HalluTracer: Hallucination Detection via Depth-Averaging Truth Signals
2026-08-17 • Computation and Language
Computation and LanguageArtificial Intelligence
AI summaryⓘ
The authors studied how big language models sometimes confidently say things that aren't true, which is a problem for important uses. They found that the models do have hidden signals about truthfulness spread throughout all the layers of their processing steps. Instead of looking at just one layer or part, the authors created HalluTracer, a method that combines information from every layer to better detect falsehoods before the model answers. This approach works better than past methods across different models and tests, showing that it’s best to gather truth clues from all layers rather than picking just one.
large language modelshallucinationtruthfulness signalswhite-box detectionlayer aggregationforward passgeometric analysishallucination benchmarksmodel interpretabilityrepresentation learning
Authors
Zhihao Guo, Zonghan Wu, Huan Huo, DaYong Ye, Junwei Zhang, Weiran Yao, Zhiwei Liu, Qingsong Wen, Yilei Shao
Abstract
Even well-aligned large language models confidently generate factually incorrect text, making hallucination a persistent reliability risk in high-stakes deployments. These models nonetheless carry linearly separable truthfulness signals in their internal representations. Existing white-box detectors, however, collapse this evidence to isolated components or a single depth, discarding discriminative information distributed across the full forward pass. We introduce HalluTracer, a detection framework that reads and aggregates truthfulness evidence across every layer of the forward pass before the model emits any answer token. A geometric analysis reveals that the per-layer signals are weakly correlated, so that simple depth averaging suppresses layer-specific noise and captures nearly all linearly accessible information. Across six open-source language models and five hallucination benchmarks, HalluTracer consistently outperforms matched white-box baselines, with gains ranging from one to fourteen points. Collectively, our work recasts hallucination detection from a layer-selection problem into a depth-aggregation problem governed by the geometric sparsity of the truthfulness signal.