arXiv Machine Learning

Attention Dispersion as a Diagnostic Signal for Hallucination in Large Language Models

The paper proposes using the temporal volatility of internal attention mechanisms—measured by an unsupervised attention dispersion metric—as a diagnostic signal for hallucinations in large language models. It demonstrates that spikes in attention entropy within intermediate layers correlate with reasoning breakdowns, and shows statistically significant AUC improvements of up to +0.076 over output-based baselines on GSM8K and MATH-500 benchmarks using the Qwen2.5 model family.

arXiv AI
Aug 5

UHP Detection: LVLMs have their Unique Hallucination Pattern in the Consistency Space

arXiv:2608. 03817v1 Announce Type: cross Abstract: Large vision--language models (LVLMs) demonstrate strong multimodal reasoning capabilities but remain prone to hallucination, where model predictions are not grounded in visual evidence.

By Amir Mohammad Ezzati, Kiyan Rezaee, Bardiya Kariminia, Mohamad Amin Yousefi, Asal Mohammadjafari Mamaqani, Behrad Samimi, Mohammad Hossein Rohban
arXiv Machine Learning
Sep 21

Gradient-Stable Attention Heads Signal LLM Correctness

The paper introduces HeadEntropy, a training‑free method that predicts the correctness of large language model (LLM) answers by measuring how stable each attention head’s pattern is to further gradient updates. By linking the trace of the softmax Jacobian to 2‑Renyi entropy, the authors show that attention spread correlates with gradient stability, enabling accurate hallucination detection without reference annotations. Across five instruction‑tuned LLMs and five diverse datasets—including medicine, multi‑hop reasoning, and mathematics—HeadEntropy achieves a 0.736 AUROC, outperforming other training‑free baselines and matching hidden‑state probes while incurring less than 1% of inference cost.

By Sophie Ostmeier, Brian Axelrod, Maya Varma, Asad Aali, Yabin Zhang, Magdalini Paschali, Sanmi Koyejo, Curtis Langlotz, Akshay Chaudhari
arXiv AI
Oct 19

Cross-Layer Attention Probing for Fine-Grained Hallucination Detection

Cross-Layer Attention Probing (CLAP) is a new activation‑probing method that examines the entire residual stream of Large Language Models (LLMs) to detect hallucinations. Experiments on five LLMs across three tasks demonstrate that CLAP outperforms baseline detectors for both greedy and temperature‑sampled outputs, enabling fine‑grained discrimination between hallucinated and non‑hallucinated responses. The approach also supports a detect‑then‑mitigate strategy that reduces hallucinations and maintains reliability even on out‑of‑distribution data.

By Malavika Suresh, Rahaf Aljundi, Ikechukwu Nkisi-Orji, Nirmalie Wiratunga
arXiv Computer Vision
Nov 5

Entropogram-guided Visual Realignment Decoding: Capturing the Rhythm of Multimodal Reasoning

The paper introduces Entropogram-guided Visual Realignment Decoding, a method that monitors entropy dynamics during multimodal large language model (MLLM) decoding to detect hallucinations. By identifying high-confidence premature convergence and abrupt entropy transitions as hallmarks of hallucination, the approach performs targeted visual re‑alignment to re‑examine salient visual evidence. Experiments on six benchmarks across nine settings show significant gains in reasoning accuracy and hallucination robustness, offering new insights into MLLM generation dynamics.

By Jiaqi Deng, Zhe Fu, Yakun Chen, Dichang Zhang, Zonghan Wu, Xiaoshui Huang, Huan Huo, Guandong Xu
arXiv AI
Oct 2

External Observers May See More Clearly: Cross-Model Span-Level Hallucination Detection in Large Language Models via Hidden State Probing

The paper proposes a hidden‑state probing method for detecting hallucinations at the span level in large language model outputs, moving beyond token‑wise binary classification. By examining layer‑wise activation patterns, the approach identifies the exact onset and continuation tokens of hallucinations, achieving higher precision‑recall AUC than random baselines despite class imbalance. Additionally, the authors introduce a cross‑model detection framework where one model observes another’s internal representations, showing that an external observer can match or surpass the generator’s own self‑detection of hallucination onsets, even when the observer is smaller.

By Kingshuk Gupta, Davide Buscaldi
arXiv Computation and Language
Oct 21

Detecting Errors in Large Language Model Responses: The Incremental Predictive Value of Hidden States

The paper investigates whether the internal hidden states of large language models can improve the detection of hallucinations in model responses. By integrating hidden states with an output‑side baseline and evaluating on both open‑ended generation and multiple‑choice tasks, the authors find that adding hidden states yields modest AUROC gains of about 0.004 and 0.039. These results suggest that while hidden states contain additional predictive information, the overall improvement over strong standalone detection methods remains limited.

By Ningzhi Li, Xiaojun Ma, Xiaofei Wang