Intrinsic Self-Correction in LLMs: Towards Explainable Prompting via Mechanistic Interpretability
This provides an interpretability-based explanation for intrinsic self-correction in LLMs, addressing a gap in understanding prompting mechanisms for researchers and practitioners, though it is incremental in building on existing mechanistic interpretability work.
The paper tackled the problem of understanding how language models refine their own outputs through prompting without external feedback, finding that self-correction prompts steer hidden representations along interpretable latent directions, such as aligning with non-toxic or toxic directions in detoxification and toxification tasks across 5 LLMs.
Intrinsic self-correction refers to the phenomenon where a language model refines its own outputs purely through prompting, without external feedback or parameter updates. While this approach improves performance across diverse tasks, its internal mechanism remains poorly understood. We analyze intrinsic self-correction from a representation-level perspective. We formalize and introduce the notion of a prompt-induced shift, which is the change in hidden representations caused by a self-correction prompt. Across 5 open-source LLMs, prompt-induced shifts in text detoxification and text toxification align with latent directions constructed from contrastive pairs. In detoxification, the shifts align with the non-toxic direction; in toxification, they align with the toxic direction. These results suggest that intrinsic self-correction functions as representation steering along interpretable latent directions, beyond what standard metrics such as task scores or model confidence capture. Our analysis offers an interpretability-based account of intrinsic self-correction and contributes to a more systematic understanding of LLM prompting.