Step Back to Leap Forward: Self-Backtracking for Boosting Reasoning of Language Models
Xiao-Wen Yang, Xuan-Yi Zhu, Wen-Da Wei, Ding-Chu Zhang, Jie-Jing Shao, Zhi Zhou, Lan-Zhe Guo, Yu-Feng Li
2025-02-10
Summary
This paper talks about a new method called self-backtracking that helps AI models think more carefully and efficiently by allowing them to go back and fix mistakes during reasoning, making them smarter and faster.
What's the problem?
AI models often struggle with complex reasoning tasks because they can't properly revisit or rethink their steps when they make mistakes. This leads to inefficient thinking, wasted effort, and a reliance on extra tools to guide them, which limits their ability to improve independently.
What's the solution?
The researchers introduced self-backtracking, a mechanism that lets AI models decide on their own when to go back and review their reasoning steps. This process is built into both the training and usage stages of the model, helping it learn from its mistakes and improve how it solves problems. The method transforms slow, careful thinking into faster and more accurate reasoning over time.
Why it matters?
This matters because it makes AI models better at handling complicated tasks like solving puzzles or answering tough questions. By improving their reasoning skills, these models can become more reliable and efficient, paving the way for smarter AI systems that require less human intervention.
Abstract
The integration of slow-thinking mechanisms into large language models (LLMs) offers a promising way toward achieving Level 2 AGI Reasoners, as exemplified by systems like OpenAI's o1. However, several significant challenges remain, including inefficient overthinking and an overreliance on auxiliary reward models. We point out that these limitations stem from LLMs' inability to internalize the search process, a key component of effective reasoning. A critical step toward addressing this issue is enabling LLMs to autonomously determine when and where to backtrack, a fundamental operation in traditional search algorithms. To this end, we propose a self-<PRE_TAG>backtracking mechanism</POST_TAG> that equips LLMs with the ability to backtrack during both training and inference. This mechanism not only enhances reasoning ability but also efficiency by transforming slow-thinking processes into fast-thinking through self-improvement. Empirical evaluations demonstrate that our proposal significantly enhances the reasoning capabilities of LLMs, achieving a performance gain of over 40 percent compared to the optimal-path supervised fine-tuning method. We believe this study introduces a novel and promising pathway for developing more advanced and robust Reasoners.