On the Intrinsic Self-Correction Capability of LLMs: Uncertainty and Latent Concept

TL;DR

This paper investigates the intrinsic self-correction ability of Large Language Models, showing it can be improved through iterative interactions, leading to stable performance by reducing model uncertainty and activating latent concepts.

Contribution

It introduces a mathematical framework and simulation to explain how self-correction converges by reducing uncertainty and activating latent concepts in LLMs.

Findings

Intrinsic self-correction improves with iterations in multi-round QA.

Iterative instructions reduce model uncertainty and calibration error.

Self-correction converges to stable performance through uncertainty reduction.

Abstract

Large Language Models (LLMs) are able to improve their responses when instructed to do so, a capability known as self-correction. When instructions provide only the task's goal without specific details about potential issues in the response, LLMs must rely on their internal knowledge to improve response quality, a process referred to as intrinsic self-correction. The empirical success of intrinsic self-correction is evident in various applications, but how and why it is effective remains unknown. In this paper, we unveil that intrinsic self-correction can be progressively improved, allowing it to approach a converged state. Our findings are verified in: (1) the scenario of multi-round question answering, by comprehensively demonstrating that intrinsic self-correction can progressively introduce performance gains through iterative interactions, ultimately converging to stable performance; and (2) the context of intrinsic self-correction for enhanced morality, in which we provide empirical evidence that iteratively applying instructions reduces model uncertainty towards convergence, which then leads to convergence of both the calibration error and self-correction performance, ultimately resulting in a stable state of intrinsic self-correction. Furthermore, we introduce a mathematical formulation and a simulation task indicating that the latent concepts activated by self-correction instructions drive the reduction of model uncertainty. Based on our experimental results and analysis of the convergence of intrinsic self-correction, we reveal its underlying mechanism: consistent injected instructions reduce model uncertainty which yields converged, improved performance.