AI 中文总结
本文提出一种融合多信号的时序方法，通过 BiGRU 序列标注检测 token 级幻觉，在 RAGTruth 上较基线提升 11 个百分点，适用于闭源模型且泛化性良好。
AI 中文摘要
现有的 token 级幻觉检测器仅基于单一信号对每个 token 独立评分，在生成模型确信出错时表现失效。本文将幻觉视为时序延伸的片段，通过序列标注进行检测：每个 token 从 33 维特征流中评分，该特征流融合了文本统计、自然语言推理（NLI）蕴含关系及语言模型惊奇度，且无需访问模型内部结构。基于这些特征的双向门控循环单元（BiGRU）在 RAGTruth（10 个随机种子）上达到 AUC 为 0.840，较独立逻辑回归基线提升 11 个百分点（p=0.002，Wilcoxon 符号秩检验）。受控分解显示，该提升主要源于时序顺序而非模型容量：证据会从置信位置传播至片段内的模糊邻域。循环、状态空间（Mamba）及注意力架构均达到约 0.845 的上限，表明瓶颈在于特征集而非模型。由于该检测器仅读取生成文本和外部信号，可应用于闭源模型，且对训练期间未见过的语言模型生成文本仍有效，AUC 下降不足 4%。
英文摘要
Token-level hallucination detectors score each token independently from a single signal, and fail exactly when the generating model is confidently wrong. This paper instead treats hallucination as a temporally extended span and detects it by 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) over these features reaches 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 within a span. The same 0.845 ceiling recurs across recurrent, state-space (Mamba), and attention architectures, locating the bottleneck in the feature set rather than the model. Because it reads only the generated text and external signals, the detector works on closed-source models, and it keeps working on text produced by language models it never saw during training, losing under 4% AUC.
Comments17 pages, 14 figures, 23 tables. Code: https://github.com/YehudaItkin/temporal-hallucination-detection
