The paper introduces TRAILS, a taxonomy of 38 mechanisms that can compromise the reliability of large language model (LLM) read‑and‑evaluate tasks in evidence synthesis, such as study screening, data extraction, risk‑of‑bias assessment, and critical appraisal. It explains how these mechanisms can cause inconsistent or hidden failures across repeated LLM runs and outlines deployment settings, testing approaches, and mitigation strategies for each. The authors argue that repeatability or accuracy alone does not guarantee trustworthy judgments, emphasizing the need for human evaluation to verify that decisions are evidence‑based, free from external influence, and reproducible.
By Carole Lunny, Ian Goodman, Stefano Berrone, Connor TA Brenna, Emma K Reid, Ivan Perez-Neri, Shailesh Kolekar, Nityanand Jain, Ashwin Kamath, Lucas HCC Santos, Yulia Otmakhova, Heba Hussein, Diana M Cespedes, Gioia Dahdal, Salmaan Kanji
arXiv:2609.32807v2 Announce Type: replace
Abstract: Clinical LLM evaluation often emphasizes answer accuracy; however, accuracy alone does not test counterfactual consistency or demographic robustnes...
By Chaitali Deb Purkayastha, Bharath Kumar Bolla, Vishnu Surya Reddy Nandi
The paper studies how on‑policy distillation (OPD) can improve answer accuracy while reducing a model’s reliance on supporting evidence. Through matched document interventions, the authors identify an "evidence bypass" phenomenon where a student model keeps an originally correct answer even after the supporting document is changed. They propose Evidence‑Paired OPD (EP‑OPD), which uses matched documents and an answer‑level margin to recover most of the counterfactual accuracy lost by standard OPD while preserving much of its original‑document accuracy gains.
By Ruoran Xu, Yilun Zhao, Ziyu Chen, Xiaoqiang kang, Arman Cohan
arXiv:2605. 03534v2 Announce Type: replace-cross Abstract: Retrieval-augmented generation (RAG) grounds answers in retrieved passages, yet relevance does not guarantee sufficiency: a topical passage may still fail to justify the answer.
By Jingxi Qiu, Zeyu Han, Cheng Huang
The paper introduces SEAC, a method that augments a fixed fact verifier with a lightweight evidence‑acquisition policy to decide when to stop or request more evidence. SEAC evaluates each claim by retrieving evidence, running the verifier, and estimating a clarification‑utility score that balances the benefit of additional evidence against its cost. Experiments on SciFact, LIAR, and FEVER show that SEAC improves mean accuracy and macro‑F1 compared to existing fact‑verification and active‑retrieval baselines.
By Yuhang Dong, Jinghong Li, Xinyue Wang, Yuanning Cui
arXiv:2606. 23915v1 Announce Type: cross Abstract: Practice often treats automatic metrics for attribution in LLM retrieval-augmented generation as interchangeable.
By Tianyu Ding, Aditya Nannapaneni, Juan Pablo De la Cruz Weinstein
arXiv:2607. 26929v1 Announce Type: cross Abstract: The same diagnostic result can support or challenge one causal claim yet fail to address another when the claims concern different populations, outcomes, estimands, pathways, or identifying assumptions.
By Weiyi Kong, Zhuoran Li
The paper presents TRAILS, a taxonomy of 38 mechanisms that can undermine the reliability of large language model (LLM) read‑and‑evaluate tasks in evidence synthesis, such as study screening and risk‑of‑bias assessment. It distinguishes failures that cause inconsistent judgments across repeated runs from those that systematically bias outcomes without obvious signs. The authors outline deployment settings, testing approaches, and mitigation strategies, and argue that repeatability or accuracy alone cannot guarantee trustworthy judgments, emphasizing the need for human evaluation of evidence grounding and process transparency.
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 presents a calibrated evidence inference model fine‑tuned from Qwen3‑4B‑Base using low‑rank adapters and a classification head to classify trial outcomes as significant increase, significant decrease, or no significant difference. Trained on 1,431 open‑access PubMed Central articles, the released model (Apache 2.0) achieves an expected calibration error of 0.017, a log loss lower than Gemma 4 31B‑it in zero‑shot settings, and a macro‑F1 of 0.925 on the test split. Accepting its top‑80 % most confident predictions yields only a 2.6 % error rate, and ablation studies show that article context influences answers while the title alone still provides useful signal.
By Johann Emmanuel Li
The study investigates how false content in training data can influence language models’ downstream decisions even without explicit triggers. By comparing models trained on misleading versus truthful documents in a controlled decision task and a real‑world bushfire case, the authors find a gap between factual answers and the decisions derived from them: correct facts do not always lead to correct decisions, and removing an injected number does not eliminate the misleading narrative. The research highlights that data poisoning can subtly alter model behavior beyond what is detectable through direct probing.
By Lin Tian, Marian-Andrei Rizoiu
The paper evaluates Laya, a small model that processes a multiple‑choice question, its options, and a text in a single 512‑token sequence. Experiments show that when the relevant evidence lies beyond the window, accuracy drops to about 26–27%, and confidence metrics fail to warn of truncation. Fine‑tuning on longer documents improves performance on visible evidence but not on cut evidence, and a larger window or a 312‑token scan can yield modest gains.
By Mohammed Nayeem Khan, Allan Kevin T