Rethinking RL for LLM Reasoning: It's Sparse Policy Selection, Not Capability Learning

Benchmark Model Rank Results
mathematical-reasoning-on-minerva-mathReasonMaxxer (Qwen3-4B)#36Accuracy: 30.5
mathematical-reasoning-on-olympiadbenchReasonMaxxer (Qwen2.5-7B)#43Accuracy: 41.1000
arithmetic-reasoning-on-gsm8kReasonMaxxer (Qwen3-4B)#56Accuracy: 91.9
mathematical-reasoning-on-math-500ReasonMaxxer (Qwen2.5-7B)#111Accuracy: 70.6