Stop When Reasoning Converges: Semantic-Preserving Early Exit for Reasoning Models

Benchmark Model Rank Results
mathematical-reasoning-on-olympiadbenchQwen3-30B-A3B-Thinking with PUMA#5Accuracy: 72.3
mathematical-reasoning-on-aime24Qwen3-30B-A3B-Thinking with PUMA#12Acc: 90.0
mathematical-reasoning-on-math-500Qwen3-30B-A3B-Thinking with PUMA#36Accuracy: 94.2
mathematical-reasoning-on-aime-2025Qwen3-30B-A3B-Thinking with PUMA#41Accuracy: 80.0
multiple-choice-question-answering-mcqa-on-gpqa-diamondQwen3-30B-A3B-Thinking with PUMA#67Accuracy: 75.8