Newest papers

Reasoning-Aware Multicalibrated Reliability Predictors for Language Models

Ekaterina Fedorova, Alessandro Morosini, Chara Podimata · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.26645 · cited by 0 works (OpenAlex) · open-access PDF

Korean

사람들은 정보에 대한 대규모 언어 모델 (LLM)에 의존하기 때문에, 모델 프레임 제공자가 개별 응답이 정확한지 또는 더 일반적으로 신뢰할 수 있는지에 대한 의미있는 예측을 전달할 필요가 있습니다. 해석의 맥락에서, 기존 작업은 대규모 응용 프로그램 관련 그룹에 대한 그러한 신뢰성 예측을 의미있게 만들기 위해 multicalibration을 사용합니다. 그러나 주제 그룹 혼자서는 개별 응답이 정확한지 또는 더 일반적으로 신뢰할 수있는지에 대한 의미있는 예측을 전달할 수 있습니다. 같은 주제 내에서 동일한 주제 및 신뢰 수준 내에서, 다른 생각 체인 (CoT) 속성을 가진 응답은 평균적으로 취소하는 반대되는 오류를 나타낼 수 있습니다. 이 작업에서, 우리는 응용 프로그램 관련 주

Machine translation (Meta M2M100); the English original follows.

Abstract

As people increasingly rely on large language models (LLMs) for information, model providers need to communicate meaningful estimates of whether individual responses are correct or, more generally, reliable. In the context of hallucinations, existing work uses multicalibration to make such reliability predictions meaningful across application-relevant groups, such as prompt topics. Topic groups alone, however, can conceal systematic errors: within the same topic and confidence level, responses with different chain-of-thought (CoT) properties may exhibit opposing errors that cancel on average. In this work, we introduce reasoning-aware multicalibration, a framework that supplements application-relevant topic groups with groups derived from observable properties of CoT traces. We highlight that, theoretically, groupings based on CoT are "internally important", i.e., groupings for which the multicalibration guarantee is not societally crucial as it is for high-risk topics, but whose inclusion in the multicalibration procedure could easily and significantly improve predictive accuracy of the final reliability predictor. Empirically, we apply our framework to hallucination prediction and unsafe-response detection across a variety of benchmarks, and find that adding CoT groups to topic-based calibration reduces Brier score in both tasks. Interestingly, we find that multicalibrating with respect to CoT groups alone (without any topic information) already improves calibration across topics, suggesting that the predictor's errors depend on how the model reasoned, rather than what it was asked about in the prompt.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Reliability (semiconductor), Variety (cybernetics), Computer science, Context (archaeology), Calibration, Machine learning

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange