Decomposing LLM Self-Correction: The Accuracy-Correction Paradox and Error Depth Hypothesis
This work addresses a critical issue for AI researchers and developers by challenging linear assumptions about model capability and self-improvement, with implications for designing self-refinement pipelines, though it is incremental in analyzing existing models.
The study tackled the problem of ineffective intrinsic self-correction in Large Language Models by decomposing it into error detection, localization, and correction, revealing an Accuracy-Correction Paradox where weaker models (e.g., GPT-3.5 with 66% accuracy) achieved 1.6x higher intrinsic correction rates (26.8% vs. 16.7%) than stronger models (e.g., DeepSeek with 94% accuracy).
Large Language Models (LLMs) are widely believed to possess self-correction capabilities, yet recent studies suggest that intrinsic self-correction--where models correct their own outputs without external feedback--remains largely ineffective. In this work, we systematically decompose self-correction into three distinct sub-capabilities: error detection, error localization, and error correction. Through cross-model experiments on GSM8K-Complex (n=500 per model, 346 total errors) with three major LLMs, we uncover a striking Accuracy-Correction Paradox: weaker models (GPT-3.5, 66% accuracy) achieve 1.6x higher intrinsic correction rates than stronger models (DeepSeek, 94% accuracy)--26.8% vs 16.7%. We propose the Error Depth Hypothesis: stronger models make fewer but deeper errors that resist self-correction. Error detection rates vary dramatically across architectures (10% to 82%), yet detection capability does not predict correction success--Claude detects only 10% of errors but corrects 29% intrinsically. Surprisingly, providing error location hints hurts all models. Our findings challenge linear assumptions about model capability and self-improvement, with important implications for the design of self-refinement pipelines.