CrEst: Credibility Estimation for Contexts in LLMs via Weak Supervision

2025/06/17 by Dyah Adila, Shuai Zhang, Adila, Dyah +7
Computer Science · Social Sciences · #Anomaly Detection Techniques and Applications #Computation and Language (cs.CL) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Misinformation and Its Impacts #Topic Modeling

paper · pdf · doi:10.48550/arxiv.2506.14912

openalex publication_date 2025/06/17 · openalex created_date 2025/10/19 · openalex updated_date 2026/07/28

Abstract

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.

Citations

• Smoothie: Label Free Language Model Routing
• CrAM: Credibility-Aware Attention Modification in LLMs for Combating Misinformation in RAG
• NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models
• Certifiably Robust RAG against Retrieval Corruption
• Not All Contexts Are Equal: Teaching LLMs Credibility-aware Generation
• RAFT: Adapting Language Model to Domain Specific RAG
• Tell Your Model Where to Attend: Post-hoc Attention Steering for LLMs
• CareerX: A Retrieval-Augmented Generation Framework for Personalized AI-Driven Career Guidance
• Toolformer: Language Models Can Teach Themselves to Use Tools
• Large Language Models Can Be Easily Distracted by Irrelevant Context
• When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories
• Atlas: Few-shot Learning with Retrieval Augmented Language Models
• SimCSE: Simple Contrastive Learning of Sentence Embeddings
• SimCSE: Simple Contrastive Learning of Sentence Embeddings
• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
• Fast and Three-rious: Speeding Up Weak Supervision with Triplet Methods
• REALM: Retrieval-Augmented Language Model Pre-Training
• Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks
• HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering
• Heterogeneous Supervision for Relation Extraction: A Representation Learning Approach
• TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for\n Reading Comprehension
• A Study of Factuality, Objectivity and Relevance: Three Desiderata in Large-Scale Information Retrieval?
• InstructRAG: Instructing Retrieval-Augmented Generation via Self-Synthesized Rationales

Related

• Not All Contexts Are Equal: Teaching LLMs Credibility-aware Generation
• Is Factuality Enhancement a Free Lunch For LLMs? Better Factuality Can Lead to Worse Context-Faithfulness
• TrustScore: Reference-Free Evaluation of LLM Response Trustworthiness
• Weakly Supervised Veracity Classification with LLM-Predicted Credibility Signals
• Calibrating LLMs with Information-Theoretic Evidential Deep Learning