CrEst: Credibility Estimation for Contexts in LLMs via Weak Supervision

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Adila, Dyah, Zhang, Shuai, Han, Boran, Min, Bonan, Wang, Yuyang
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866908411294121984
author Adila, Dyah
Zhang, Shuai
Han, Boran
Min, Bonan
Wang, Yuyang
author_facet Adila, Dyah
Zhang, Shuai
Han, Boran
Min, Bonan
Wang, Yuyang
contents The integration of contextual information has significantly enhanced the performance of large language models (LLMs) on knowledge-intensive tasks. However, existing methods often overlook a critical challenge: the credibility of context documents can vary widely, potentially leading to the propagation of unreliable information. In this paper, we introduce CrEst, a novel weakly supervised framework for assessing the credibility of context documents during LLM inference--without requiring manual annotations. Our approach is grounded in the insight that credible documents tend to exhibit higher semantic coherence with other credible documents, enabling automated credibility estimation through inter-document agreement. To incorporate credibility into LLM inference, we propose two integration strategies: a black-box approach for models without access to internal weights or activations, and a white-box method that directly modifies attention mechanisms. Extensive experiments across three model architectures and five datasets demonstrate that CrEst consistently outperforms strong baselines, achieving up to a 26.86% improvement in accuracy and a 3.49% increase in F1 score. Further analysis shows that CrEst maintains robust performance even under high-noise conditions.
format Preprint
id arxiv_https___arxiv_org_abs_2506_14912
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle CrEst: Credibility Estimation for Contexts in LLMs via Weak Supervision
Adila, Dyah
Zhang, Shuai
Han, Boran
Min, Bonan
Wang, Yuyang
Computation and Language
Machine Learning
The integration of contextual information has significantly enhanced the performance of large language models (LLMs) on knowledge-intensive tasks. However, existing methods often overlook a critical challenge: the credibility of context documents can vary widely, potentially leading to the propagation of unreliable information. In this paper, we introduce CrEst, a novel weakly supervised framework for assessing the credibility of context documents during LLM inference--without requiring manual annotations. Our approach is grounded in the insight that credible documents tend to exhibit higher semantic coherence with other credible documents, enabling automated credibility estimation through inter-document agreement. To incorporate credibility into LLM inference, we propose two integration strategies: a black-box approach for models without access to internal weights or activations, and a white-box method that directly modifies attention mechanisms. Extensive experiments across three model architectures and five datasets demonstrate that CrEst consistently outperforms strong baselines, achieving up to a 26.86% improvement in accuracy and a 3.49% increase in F1 score. Further analysis shows that CrEst maintains robust performance even under high-noise conditions.
title CrEst: Credibility Estimation for Contexts in LLMs via Weak Supervision
topic Computation and Language
Machine Learning
url https://arxiv.org/abs/2506.14912