Beyond Output Critique: Self-Correction via Task Distillation

TL;DR

This work reframes self-correction as task understanding before solution refinement. By introducing Self-Thought, which distills input problems into structured templates prior to solving, the approach yields more robust and generalizable corrections than traditional output-focused edits. Distil-Thought extends this benefit to smaller models by transferring abstractions generated by larger models, enabling reliable self-correction without fine-tuning or external verifiers. Empirical results across diverse reasoning tasks and model scales demonstrate consistent accuracy gains, particularly in early iterations, and show that task abstractions serve as effective control representations for downstream reasoning. Overall, the framework offers a scalable path to reliable self-correcting LLMs across both large and small models.

Abstract

Large language models (LLMs) have shown promising self-correction abilities, where iterative refinement improves the quality of generated responses. However, most existing approaches operate at the level of output critique, patching surface errors while often failing to correct deeper reasoning flaws. We propose SELF-THOUGHT, a framework that introduces an intermediate step of task abstraction before solution refinement. Given an input and an initial response, the model first distills the task into a structured template that captures key variables, constraints, and problem structure. This abstraction then guides solution instantiation, grounding subsequent responses in a clearer understanding of the task and reducing error propagation. Crucially, we show that these abstractions can be transferred across models: templates generated by larger models can serve as structured guides for smaller LLMs, which typically struggle with intrinsic self-correction. By reusing distilled task structures, smaller models achieve more reliable refinements without heavy fine-tuning or reliance on external verifiers. Experiments across diverse reasoning tasks demonstrate that SELF-THOUGHT improves accuracy, robustness, and generalization for both large and small models, offering a scalable path toward more reliable self-correcting language systems.

Figures (8)

• Figure 1: We present an example trace of Self-Thought self-correcting on a sample from AIME 2025 using GPT-4o-Mini. The initial answer is simplified for clarity, and the full response is provided in Table \ref{['tab:full-answer']} in the Appendix. This initial response contains logical reasoning errors, incomplete calculations, and an incorrect final result. By applying task abstraction, Self-Thought successfully identifies and corrects these mistakes.
• Figure 2: Accuracy over iterations with self-correction methods across models. Top row show results on AIME 2024 using large models, while Bottom row show results on AIME 2024 (subfigures e and f) and AIME 2025 (subfigures g and h). Please refer to Figures \ref{['fig:app:iter-large']} and \ref{['fig:app:iter-small']} in the Appendix for the iteration effect plots of other tasks.
• Figure 3: Comparison of Self-Thought and Distil-Thought with the Self-Consistency on o3-mini and Qwen-2.5-7B. See Figure \ref{['fig:app:self-consistency']} in Appendix \ref{['sec:app:add-self-consistency']} for the results on other models.
• Figure 4: Accuracy over iterations on (Right) AIME 2024 and (Left) AIME 2025.
• Figure 5: Performance comparison of Self-Thought and ablated variants for AIME 2024 and AIME 2025.