Self-Correction Bench: Uncovering and Addressing the Self-Correction Blind Spot in Large Language Models

TL;DR

This paper identifies a systematic failure in large language models where they cannot correct their own errors, introduces an evaluation framework to measure this issue, and proposes a simple prompt modification to significantly reduce the problem.

Contribution

The paper uncovers the Self-Correction Blind Spot in LLMs, introduces Self-Correction Bench for evaluation, and demonstrates a prompt-based method to mitigate the issue.

Findings

Average 64.5% blind spot rate across models

Appending a 'Wait' prompt reduces blind spots by 89.3%

Training data influences error correction capabilities

Abstract

Although large language models (LLMs) have transformed AI, they still make mistakes and can explore unproductive reasoning paths. Self-correction capability is essential for deploying LLMs in safety-critical applications. We uncover a systematic failure: LLMs cannot correct errors in their own outputs while successfully correcting identical errors from external sources - a limitation we term the Self-Correction Blind Spot. To study this phenomenon, we introduce Self-Correction Bench, an evaluation framework to measure this phenomenon through controlled error injection at three complexity levels. Testing 14 open-source non-reasoning models, we find an average 64.5% blind spot rate. We provide multiple lines of evidence suggesting this limitation may be influenced by training data: human demonstrations rarely include error-correction sequences (favoring error-free responses), whereas reinforcement learning (RL) trained models learn error correction via outcome feedback. Remarkably, appending a minimal "Wait" prompt activates a 89.3% reduction in blind spots, suggesting dormant capabilities that require triggering. Our work highlights a critical limitation potentially influenced by training distribution and offers a practical approach to enhance LLM reliability and trustworthiness - vital for safety-critical domains.