The paper demonstrates that open‑ended Theory‑of‑Mind trackers can produce valid beliefs that are absent from finite reference sets, and that treating unmatched outputs as false can reverse model‑selection rankings. By recoding references for 259 beliefs, the authors show a dramatic drop in weighted prevalence and a reversal of strictly proper Brier risk, with similar distortions observed in a 301‑question NQ‑open DPR‑BERT pipeline. The study further reveals that 90‑96% of audited unmatched beliefs are literally true, and introduces a TriSource‑Restore method that anchors reference labels to a probability‑sampled human pilot to restore calibration and ranking integrity.
By Zhexi Feng, Wuxi Chen, Bingrui Zhang
The paper audits a 366‑day autobiographical book generated by a large language model (LLM) against an independent verification corpus. Using a four‑level rubric, 354 of the 366 days (96.7%) failed verification, with only 12 days containing corroborated scenes and 19 days containing actively contradicted claims. Regenerating the same days with current models yielded 100% verification failure, while grounding the generation in the subject’s own corpus improved the rate to 83.3% but still left substantial residual failure.
By Heather Renze
The paper revises earlier claims about hallucination rates in multi‑agent LLM systems, noting that a judge that flags contradictions can mistakenly label correct reports of teammate errors as hallucinations. Using a pre‑registered analysis plan, the authors re‑score outputs from a harness with Llama and Claude agents, finding that the broadcast effect on hallucination rates is negligible and that most hallucinations stem from teammates’ wrong values. The study also reports that the effect varies across judge‑prompt combinations and that the data are limited to a single harness and 13 hand‑written scenarios.
By Carson Rodrigues
arXiv:2609.25447v1 Announce Type: new
Abstract: We study what LLMs do when a user applies pressure in an uncomfortable situation: a user insists, begs, flatters or grieves, and the model gives up a c...
By Tapan Parikh
arXiv:2610.20417v1 Announce Type: new
Abstract: Audits of large language model (LLM) extraction from full-length documents report severe unreliability, leaving two questions open: what the failures a...
By Jiahong Qian
arXiv:2606. 10315v1 Announce Type: cross Abstract: LLM-as-judge is the default instrument for evaluating conversational agents, yet its reliability is almost always reported as agreement with human ratings, not recall of real defects.
By Sawyer Zhang, Alexander Wang, Sophie Lei
The study audits AI "humanizers"—tools that rewrite machine‑generated text to evade detection—by checking whether factual units such as numbers, dates, names, and quoted spans survive each rewrite. Using a rule‑based checker validated by AI annotators, the authors found that in the 2026 public leaderboard, 35–65% of documents had at least one factual unit altered or dropped, with quoted spans most affected. An open 12B humanizer retained most numbers and names but reworded 67–78% of quotes, and a zero‑shot detector (Jev) could distinguish human originals from humanized rewrites with an AUROC of 0.864, outperforming several existing detectors.
By Ziyuan Wang, Genliang Zhu, Xiaobo Chen
arXiv:2608.31016v1 Announce Type: cross
Abstract: Ambient AI scribes draft clinical notes, and published audits find their dominant error is omission: information the encounter established that the n...
By Sebastian Fox, Luke Markham, Ryan Lail, Michael Karotsieris
The paper proposes a training‑free detector that uses sentence‑level context sensitivity to identify unsupported content in retrieval‑augmented generation (RAG) answers. By re‑scoring each sentence with full context, no context, and each chunk removed, the method flags the chunk whose removal most reduces a sentence’s likelihood as the likely source. Evaluated on RAGTruth, TofuEval, and RAGBench, the detector outperforms answer‑level faithfulness scores, achieving AUCs up to 0.745 and matching per‑chunk fact‑checkers while using only a fraction of the compute required by large‑language‑model judges.
By Mohamed Aly Bouke
The paper argues that relying solely on overall accuracy or F1 scores to validate automatic scorers in LLM evaluation can conceal prompt‑specific errors that distort comparative results. By auditing six models on 180 invented‑entity questions across six prompts, the authors show that a scorer’s overall F1 of .916 masks substantial variation—up to 33.3%—in error rates depending on the prompt. In a second audit, differences in measured prompt effects dropped dramatically when multiple raters were used, highlighting the need to validate the exact comparison rather than the scorer alone.
By Ancil Crayton
The paper introduces EmoNet‑Face‑HQ, a fine‑grained emotion recognition benchmark that uses generated portraits and a 40‑category taxonomy to evaluate vision‑language models (VLMs). It finds that VLMs perform poorly when asked to generate responses but can match or surpass a fine‑tuned model (Empathic‑Insight‑Face) when their logits are read directly as binary queries. The study shows that the benchmark’s difficulty lies in the readout process rather than in perception, and that graded probability outputs yield better performance than simple yes/no questions.
By Tobias Hallmen, Fabian Deuser, Robin-Nico Kampa, Norbert Oswald, Elisabeth Andr\'e
Almost every paper on LLM jailbreaks and prompt injection reports an attack-success rate (ASR), and that number is assigned not by people but by an automated judge: either a safety classifier trained for the task, or a general chat model prompted to grade. The judge is rarely checked.