LLM Self-Correction Paradox: Weaker Models Outperform in Error Recovery

This research highlights a critical flaw in the assumption that stronger LLMs are inherently better at self-correction, revealing a counterintuitive relationship between accuracy and correction rate. The Error Depth Hypothesis offers a plausible explanation, suggesting that advanced models generate more complex errors that are harder to rectify internally. This has significant implications for designing effective self-refinement strategies and understanding the limitations of current LLM architectures.

• •Weaker LLMs exhibit higher intrinsic self-correction rates than stronger LLMs.
• •Error detection capability does not directly correlate with correction success.
• •Providing error location hints negatively impacts self-correction performance.