RISE: Recursive Improvement via Self-Extrapolating Policy Distillation

Benchmark Model Rank Results
mathematical-reasoning-on-aime-2025RISE (logit,Qwen3-8B,beta0=1.5)–Accuracy: 47.9
mathematical-reasoning-on-aime24RISE (logit,Qwen3-8B,cosineschedule)–Acc: 58.5
mathematical-reasoning-on-math-500RISE (weight,Qwen3-8B,3-seedmean)–Accuracy: 84.9
mathematical-reasoning-on-minerva-mathRISE (weight,Qwen3-8B)–Accuracy: 32.6
mathematical-reasoning-on-olympiadbenchRISE (weight,Qwen3-8B,3-seedmean)–Accuracy: 64.4
multi-task-language-understanding-on-mmlu-proRISE (logit,Qwen3-8B)–Accuracy: 74.3
multiple-choice-question-answering-mcqa-on-gpqa-diamondRISE (weight,Qwen3-8B)–Accuracy: 59.0