Abstract:Large Language Models (LLMs) are often augmented with external information as contexts, but this external information can sometimes be inaccurate or even intentionally misleading. 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 benchmark this capability, we evaluate LLMs across several QA datasets, including a newly created dataset called RedditQA 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-access 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. Our findings highlight promising avenues for improving situated faithfulness in LLMs. The data and code are released.

Calibrating Large Language Models Using Their Generations Only

Calibrating Long-form Generations from Large Language Models

The Calibration Gap between Model and Human Confidence in Large Language Models

Calibrating Large Language Models with Sample Consistency

Graph-based Confidence Calibration for Large Language Models

Multicalibration for Confidence Scoring in LLMs

Large Language Models Must Be Taught to Know What They Don't Know

Calibrating the Confidence of Large Language Models by Eliciting Fidelity

Calibrated Large Language Models for Binary Question Answering

Linguistic Calibration of Long-Form Generations

A Survey of Confidence Estimation and Calibration in Large Language Models

Atomic Calibration of LLMs in Long-Form Generations

Large Language Model Confidence Estimation via Black-Box Access

Calibrating LLM-Based Evaluator

Confidence Calibration and Rationalization for LLMs via Multi-Agent Deliberation

Enhancing Healthcare LLM Trust with Atypical Presentations Recalibration

Calibrating Verbalized Probabilities for Large Language Models

Enhancing Large Language Models' Situated Faithfulness to External Contexts

Generative Calibration for In-context Learning

A Survey of Calibration Process for Black-Box LLMs