The paper investigates how the topology of attention graphs can differentiate hallucinated from non-hallucinated responses in large language models. By analyzing Forman-Ricci curvature, the authors identify structural bottlenecks and develop a method that captures both semi-local and global information-flow characteristics of attention heads associated with hallucinations. Extensive evaluation across multiple LLMs and benchmarks shows that this single-pass approach consistently outperforms existing attention-based and multi-response baselines, while also revealing that impaired context sharing—such as over-reliance on self-attention and information over-squashing—correlates strongly with hallucination occurrences.
arXiv:2604. 10697v2 Announce Type: replace-cross Abstract: Large language models frequently exhibit hallucinations: fluent and confident outputs that are factually incorrect or unsupported by the input context.
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.
Optimal transport (OT) has been shown to detect hallucinations in neural machine translation (NMT) by measuring the geometric distance between cross-attention distributions and a reference distribution, without any supervision. We extend this analysis to all six decoder layers of the Fairseq DE-EN model ($N=3{,}414$), showing that Wass-to-Unif and Wass-to-Data are complementary detectors specialised across hallucination types, that detection is concentrated in layers L1--L4 with L5 anti-predictive for subtler types, and that hallucinated translations lack the exploratory attention phase present in correct translations from the first decoding step.
arXiv:2509. 24770v2 Announce Type: replace Abstract: Large Language Models (LLMs) often generate incorrect or unsupported content, known as hallucinations.
arXiv:2606. 31054v1 Announce Type: cross Abstract: Multimodal Large Language Models (MLLMs) are critically hampered by hallucination, generating content inconsistent with the provided image.
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.
arXiv:2608. 10835v1 Announce Type: cross Abstract: Large Vision-Language Models (LVLMs) achieve impressive visual reasoning and dialogue capabilities, yet frequently hallucinate content unsupported by the visual input.
The paper introduces InnerExpert, a method that uses Mixture-of-Experts (MoE) architecture signals—such as router entropy, expert disagreement, and usage patterns—to detect hallucinations at the token level in Large Language Models. By combining these MoE-specific signals with standard transformer features into compact per-token vectors, InnerExpert trains a lightweight detector using an LLM-as-a-judge pipeline, enabling continuous updates without manual labeling. Experiments across five datasets and two MoE architectures show that InnerExpert outperforms existing methods, achieving up to 0.91 answer-level and 0.76 token-level AUROC with only a single forward pass.
arXiv:2608. 07302v1 Announce Type: cross Abstract: Large Vision-Language Models (LVLMs) often suffer from object hallucination, generating objects that are absent from the image.
The paper investigates whether recent attention‑mechanism improvements—specifically gated attention, Kimi K3, Kimi Delta Attention, and Attention Residuals—effectively eliminate the attention‑sink problem when scaling language models to a one‑million‑token context window. Using a new diagnostic suite called SinkProbe, the authors evaluate sink mass, massive activation, position‑resolved recall, and the recency gap across four small models that vary only in token mixing and depth. Their findings show that the training objective, rather than the architecture, drives the emergence of attention sinks; gating did not replicate its previously reported benefits at the larger scale, and sink mass, activations, and positional bias behaved independently.
arXiv:2609.37263v1 Announce Type: new Abstract: While Large Vision-Language Models (LVLMs) achieve remarkable success, hallucinations remain a significant barrier to their reliable deployment. Recent...
The paper introduces InnerExpert, a method that uses Mixture-of-Experts (MoE) internal signals—such as router entropy, expert disagreement, and usage patterns—to detect hallucinations at the token level in large language models. By combining these MoE-specific signals with standard transformer features into compact per-token vectors, InnerExpert trains a lightweight detector using an LLM-as-a-judge pipeline, enabling continuous updates without manual labeling. Experiments across five datasets and two MoE architectures show that InnerExpert outperforms existing methods, achieving up to 0.91 answer-level and 0.76 token-level AUROC with only a single forward pass.