Can LLMs Learn from Previous Mistakes? Investigating LLMs' Errors to Boost for Reasoning

TL;DR

The paper investigates whether LLMs can improve reasoning by learning from their own mistakes. It introduces CoTErrorSet, a large-scale benchmark of correct and incorrect chain-of-thought rationales across diverse tasks, and proposes two learning-from-mistakes paradigms: self-rethinking (iterative introspection during inference) and mistake tuning (finetuning on mixed correct/incorrect rationales with prefixes). Across arithmetic and commonsense reasoning, these methods yield consistent accuracy gains and demonstrate cost-effective benefits compared to traditional gold-only fine-tuning. The work also analyzes the spectrum of reasoning errors to guide future improvements, and offers practical pathways for leveraging model mistakes to boost reasoning capabilities. The dataset and methods have broad implications for scalable, error-informed reasoning enhancements in LLMs.

Abstract

Recent works have shown the benefits to LLMs from fine-tuning golden-standard Chain-of-Thought (CoT) rationales or using them as correct examples in few-shot prompting. While humans can indeed imitate correct examples, learning from our mistakes is another vital aspect of human cognition. Hence, a question naturally arises: \textit{can LLMs learn and benefit from their mistakes, especially for their reasoning? } This study investigates this problem from both the prompting and model-tuning perspectives. We begin by introducing \textsc{CoTErrorSet}, a new benchmark with 609,432 questions, each designed with both correct and error references, and demonstrating the types and reasons for making such mistakes. To explore the effectiveness of those mistakes, we design two methods: (1) \textbf{Self-rethinking} prompting guides LLMs to rethink whether they have made similar previous mistakes; and (2) \textbf{Mistake tuning} involves finetuning models in both correct and incorrect reasoning domains, rather than only tuning models to learn ground truth in traditional methodology. We conduct a series of experiments to prove LLMs can obtain benefits from mistakes in both directions. Our two methods offer potentially cost-effective strategies by leveraging errors to enhance reasoning capabilities, which costs significantly less than creating meticulously hand-crafted golden references. We ultimately make a thorough analysis of the reasons behind LLMs' errors, which provides directions that future research needs to overcome. \textsc{CoTErrorSet} will be published soon on \texttt{\url{https://github.com/YookiTong/Learn-from-Mistakes-CotErrorSet}}.

Figures (5)

• Figure 1: The overview pipeline of our work includes (1). Mistake collection and analysis (Section \ref{['A Novel Dataset']}). (2) Two novel methods to instruct LLMs to learn from mistakes(Section \ref{['Our Methodology: Self-rethinking']} and Section \ref{['Our Methodology: Mistake Tuning']}).
• Figure 2: The pipeline to construct CoTErrorSet. By providing PaLM2 with the correct reference and the incorrect response generated by itself, we prompt it to introspect and grasp the underlying reasons for its errors.
• Figure 3: Our pipeline for clustering PaLM2's mistakes.
• Figure 4: Accuracy of different re-thinking iterations($k$). As the value of $k$ increases, the overall prediction accuracy improves.
• Figure 5: PaLM2's error type distribution in the commonsense and arithmetic reasoning task.