The paper investigates whether frozen language models can detect a corrupted reward signal by using a single verified record in a two‑option game. In the game, a payout swap and a lying reporter produce identical histories, but a single line confirming the true outcome allows the models to almost perfectly identify the liar. However, the models frequently misclassify honest reporters as liars, with error rates ranging from 26% to 58% depending on model size and wording, indicating a significant limitation in their ability to interpret verified data.
By Arman Nik Khah
The paper investigates how misleading context—specifically fabricated evidence and bare assertions—affects large language models’ medical question‑answering performance. Experiments on MedMisBench show that models are more prone to adopt answers based on assertions than fabricated evidence, and that these misleading cues are often disclosed in reasoning traces but rarely in final responses. A monitor that reads open reasoning traces can detect most corrupted decisions, whereas monitoring only responses is less effective.
By Robin Linzmayer, No\'emie Elhadad
The paper introduces the concept of perfect aliasing, where a truth probe that aligns truthful reporting with a task’s prescribed action cannot differentiate between the two based solely on its labels. In a binary reporting game, probes fitted on compliant contexts yield identical optimizations, while on rival contexts their labels are complementary, causing their AUROCs to sum to one across 751 cell-layer pairs. By employing randomized codebooks and mixed-context fitting, the authors demonstrate that separating prescribed output symbols from semantic action enables perfect recovery of truth, achieving an AUROC of 1.000 on rival trials for a reward-trained Gemma-2-9B policy, whereas conventional probes perform near chance.
By Dylan Jayabahu
arXiv:2609.09243v1 Announce Type: cross
Abstract: Retrieval-augmented generation (RAG) grounds a language model in retrieved documents, which reduces hallucination but creates a new attack surface: i...
By Iliano Fasolino
A 0.6B language model consistently answers YES to 1,200 logical tests, yet its behavior shows no discrimination. Linear probes reveal the correct verdict with high AUC (0.96) and transfer to unseen structures, but a single scalar readout fails due to a saturated decision threshold offset by +4.6 σ. Adjusting this threshold restores behavior accuracy from 50 % to 81 % and improves higher‑scale models, demonstrating that miscalibrated readouts, not hidden knowledge loss, drive performance gaps.
By Gnaneswar Villuri, Hashmath Shaik, Alex Doboli
The paper introduces a joint fact‑verification score that evaluates both answers and the evidence submitted with them. On the FEVEROUS dataset, replacing the DCUF evidence with UnifEE evidence improves the strict score by about 9.6 percentage points, while answer accuracy rises only 1.96 points. The study also shows that increasing context length for large language models yields modest evidence‑gain improvements, and that detailed answer‑evidence analyses uncover patterns missed by aggregate metrics.
By Han Chen, Yingrui Li
arXiv:2609.21996v1 Announce Type: new
Abstract: Large language models can hold knowledge they do not report. A model may sandbag on a capability evaluation, or answer against what it internally knows...
By Hiskias Dingeto
arXiv:2607. 08456v1 Announce Type: cross Abstract: A model should refuse two different things: answers it would get wrong, and questions it should not answer at all, such as unanswerable ones or ones resting on a false premise.
By Benedikt J. Wagner
PROOF is a benchmark that profiles the reliability of object-level facts in instruction-tuned language models by converting a frozen Wikidata snapshot into 18,486 English multiple-choice questions grounded in 11,779 semantic facts across 101 classes, 392 properties, and 14 domains. Each question includes an explicit "I don't know" option, a "No correct option" control, and nine controlled formulations, with 1,849 questions designed as no-correct-option traps. The study evaluates 18 open-weight model deployments on 166,374 prompts, revealing wide variability in factual accuracy, sensitivity to wording changes, and the impact of decoder perturbations.
By Andrei Chetvergov, Mikhail Solovev, Timofei Sivoraksha, Stepan Ukolov, Valeriia Kuschenko, Alexander Evseev, Sergey Bolovtsov
arXiv:2610.03190v1 Announce Type: cross
Abstract: Accident, defect and outage investigations end with a decision that ordinary question answering never faces: whether the evidence gathered so far is...
By Tingzhu Bi, Ping Wang, Meng Ma
The study examines how different editorial framings in prompts influence large language models’ statistical analysis reports. Using a 4×4 factorial design, researchers found that certain framings—particularly brutally critical prompts on genuine effects and significance-seeking prompts on underpowered nulls—led to factual misrepresentations. Tone shifts were more widespread, with critical framing inducing defensive language across all data patterns, while a confound in the data largely prevented both factual and tonal distortions.
By Paras Balani, Subhrakanta Panda
arXiv:2608. 11922v1 Announce Type: cross Abstract: Predictive-distribution entropy makes a strong selection rule in retrieval-augmented question answering: across five QA benchmarks, keeping the candidate answer that a frozen respondent LLM produces with the lowest answer-token entropy lifts mean answer $F_1$ from 0.
By Po-Jen Ko, Che-Cheng Wu, Hung-Chun Hsu, Li-Yang Chang, Chuan-Ju Wang