To Trust or Not to Trust? Enhancing Large Language Models' Situated Faithfulness to External Contexts

TL;DR

This paper investigates how large language models can better calibrate their trust in external contexts, especially when faced with inaccurate information, by proposing and evaluating two confidence reasoning approaches.

Contribution

It introduces the concept of situated faithfulness and proposes two novel methods, SCR and RCR, to improve LLMs' ability to assess external information reliability.

Findings

SCR outperforms RCR on strong reasoning models like GPT-4.

RCR outperforms SCR on smaller models like Llama-3-8B.

Fine-tuning SCR with CR-DPO improves performance across datasets.

Abstract

Large Language Models (LLMs) are often augmented with external contexts, such as those used in retrieval-augmented generation (RAG). However, these contexts can be inaccurate or intentionally misleading, leading to conflicts with the model's internal knowledge. We argue that robust LLMs should demonstrate situated faithfulness, dynamically calibrating their trust in external information based on their confidence in the internal knowledge and the external context to resolve knowledge conflicts. To benchmark this capability, we evaluate LLMs across several QA datasets, including a newly created dataset featuring in-the-wild incorrect contexts sourced from Reddit posts. We show that when provided with both correct and incorrect contexts, both open-source and proprietary models tend to overly rely on external information, regardless of its factual accuracy. To enhance situated faithfulness, we propose two approaches: Self-Guided Confidence Reasoning (SCR) and Rule-Based Confidence Reasoning (RCR). SCR enables models to self-assess the confidence of external information relative to their own internal knowledge to produce the most accurate answer. RCR, in contrast, extracts explicit confidence signals from the LLM and determines the final answer using predefined rules. Our results show that for LLMs with strong reasoning capabilities, such as GPT-4o and GPT-4o mini, SCR outperforms RCR, achieving improvements of up to 24.2% over a direct input augmentation baseline. Conversely, for a smaller model like Llama-3-8B, RCR outperforms SCR. Fine-tuning SCR with our proposed Confidence Reasoning Direct Preference Optimization (CR-DPO) method improves performance on both seen and unseen datasets, yielding an average improvement of 8.9% on Llama-3-8B. In addition to quantitative results, we offer insights into the relative strengths of SCR and RCR.