Can LLMs Correct Themselves? A Benchmark of Self-Correction in LLMs

TL;DR

The paper tackles whether LLMs can truly correct their own outputs through a structured benchmark. CorrectBench categorizes self-correction into intrinsic, external, and fine-tuned methods and evaluates them on three tasks (commonsense, math, code) with two LLM types. Key findings show substantial accuracy gains from self-correction, especially on complex tasks, but significant efficiency costs, and that simple chain-of-thought baselines can be competitive. The work lays a foundation for adaptive, cost-aware self-correction and invites further research into dynamic selection and practical deployment.

Abstract

Self-correction of large language models (LLMs) emerges as a critical component for enhancing their reasoning performance. Although various self-correction methods have been proposed, a comprehensive evaluation of these methods remains largely unexplored, and the question of whether LLMs can truly correct themselves is a matter of significant interest and concern. In this study, we introduce CorrectBench, a benchmark developed to evaluate the effectiveness of self-correction strategies, including intrinsic, external, and fine-tuned approaches, across three tasks: commonsense reasoning, mathematical reasoning, and code generation. Our findings reveal that: 1) Self-correction methods can improve accuracy, especially for complex reasoning tasks; 2) Mixing different self-correction strategies yields further improvements, though it reduces efficiency; 3) Reasoning LLMs (e.g., DeepSeek-R1) have limited optimization under additional self-correction methods and have high time costs. Interestingly, a comparatively simple chain-of-thought (CoT) baseline demonstrates competitive accuracy and efficiency. These results underscore the potential of self-correction to enhance LLM's reasoning performance while highlighting the ongoing challenge of improving their efficiency. Consequently, we advocate for further research focused on optimizing the balance between reasoning capabilities and operational efficiency. Project Page: https://correctbench.github.io/

Figures (25)

• Figure 1: Comparison of different LLMs across various self-correction types and task scenarios.
• Figure 2: An overview of the CorrectBench framework.
• Figure 3: Mixture of different correction methods.
• Figure 4: Average performance improvements achieved by S1 across multiple LLMs.
• Figure 5: Average performance improvements achieved by S2 across multiple LLMs.