EvoThink is a framework designed to address the issue of overthinking in Large Reasoning Models (LRMs). LRMs often suffer from excessive reasoning due to redundant verification steps, and existing methods, such as fast-slow thinking switching and reasoning trajectory compression, fail to finely distinguish between beneficial and redundant steps, potentially impairing reasoning capability in the pursuit of efficiency.
The core of EvoThink consists of two key components:
- Self-Pruning Training (SPT): An unsupervised method that iteratively prunes redundant reasoning steps and self-trains on concise trajectories to enhance reasoning efficiency.
- Aha-Moment Preference Optimization (AMPO): Inspired by genetic algorithms, this component identifies valuable failed reasoning attempts, synthesizes from-wrong-to-right aha-moment data, and optimizes the model to internalize this reasoning pattern.
Extensive evaluations across mathematical reasoning and code generation benchmarks demonstrate that EvoThink not only significantly reduces inference-time token usage but also improves the reasoning capability of LRMs.
Blogger's Review: The self-pruning and aha-moment optimization mechanisms proposed by EvoThink offer a novel approach to efficient reasoning in large models. By effectively reducing redundant steps, it enhances the model's overall reasoning capability, making it a significant contribution to both practical applications and theoretical research.