Towards Fully Exploiting LLM Internal States to Enhance Knowledge Boundary Perception

TL;DR

This paper investigates how large language models can better perceive their knowledge boundaries by leveraging internal states, improving confidence estimation, efficiency, and risk control through novel calibration methods.

Contribution

It introduces a confidence calibration method ($C^3$) that enhances LLMs' ability to recognize knowledge gaps and improves their reliability in critical tasks.

Findings

LLMs show significant pre-response confidence perception.

Post-generation perception further refines confidence estimates.

The $C^3$ method increases unknown perception rate by over 5%.

Abstract

Large language models (LLMs) exhibit impressive performance across diverse tasks but often struggle to accurately gauge their knowledge boundaries, leading to confident yet incorrect responses. This paper explores leveraging LLMs' internal states to enhance their perception of knowledge boundaries from efficiency and risk perspectives. We investigate whether LLMs can estimate their confidence using internal states before response generation, potentially saving computational resources. Our experiments on datasets like Natural Questions, HotpotQA, and MMLU reveal that LLMs demonstrate significant pre-generation perception, which is further refined post-generation, with perception gaps remaining stable across varying conditions. To mitigate risks in critical domains, we introduce Confidence Consistency-based Calibration ($C^3$), which assesses confidence consistency through question reformulation. $C^3$ significantly improves LLMs' ability to recognize their knowledge gaps, enhancing the unknown perception rate by 5.6% on NQ and 4.9% on HotpotQA. Our findings suggest that pre-generation confidence estimation can optimize efficiency, while $C^3$ effectively controls output risks, advancing the reliability of LLMs in practical applications.