ResRL: Boosting LLM Reasoning via Negative Sample Projection Residual Reinforcement Learning

Benchmark Model Rank Results
mathematical-reasoning-on-minerva-mathResRL (Qwen3-8B)#15Accuracy: 46.0
mathematical-reasoning-on-math-500ResRL (Qwen3-8B)#44Accuracy: 92.7
mathematical-reasoning-on-aime24ResRL (Qwen3-8B)#81Acc: 50.8
mathematical-reasoning-on-aime-2025ResRL (Qwen3-8B)#96Accuracy: 41.1