EvoThink: Evolving Thinking in Large Reasoning Models via Self-Pruning and Aha-Moment Preference Optimization
Merged summary
TL;DR - EvoThink is a training framework that removes redundant reasoning steps while teaching large reasoning models to convert useful failed attempts into better solution paths. It reportedly reduces inference token usage while improving performance on math and code-generation benchmarks.
- Self-Pruning Training iteratively removes redundant verification steps and self-trains on the resulting concise trajectories.
- Aha-Moment Preference Optimization identifies valuable failed attempts and synthesizes wrong-to-right reasoning examples.
- The approach targets efficiency without sacrificing reasoning quality.
- Evaluations cover mathematical reasoning and code generation.
Sources (1)
EvoThink: Evolving Thinking in Large Reasoning Models via Self-Pruning and Aha-Moment Preference Optimization
TL;DR - EvoThink is a training framework that removes redundant reasoning steps while teaching large reasoning models to convert useful failed attempts into better solution paths. It reportedly reduces inference token usage while improving performance on math and code-generation benchmarks.
- Self-Pruning Training iteratively removes redundant verification steps and self-trains on the resulting concise trajectories.
- Aha-Moment Preference Optimization identifies valuable failed attempts and synthesizes wrong-to-right reasoning examples.
- The approach targets efficiency without sacrificing reasoning quality.
- Evaluations cover mathematical reasoning and code generation.