Tag
This paper introduces AdaPrefix-GRPO, a method that adaptively controls the length of correct solution prefixes provided to a model during GRPO training, maintaining a 50% success rate to maximize gradient signal. It significantly improves accuracy on hard math reasoning problems while reducing computational cost.
A method that dynamically allocates compute budget to hard problems using Qwen-35B-A3B achieves performance near GPT-5.4-xHigh on the HLE benchmark.