FlowBalance: Verifier-Grounded Self-Improvement from On-Policy Reasoning Experience

Benchmark Model Rank Results
mathematical-reasoning-on-minerva-mathQwen3-8B + FlowBalance#6Accuracy: 53.68
mathematical-reasoning-on-olympiadbenchQwen3-8B + FlowBalance#13Accuracy: 66.85
mathematical-reasoning-on-hmmt-feb-2025Qwen3-8B + FlowBalance#22Accuracy: 34.67
mathematical-reasoning-on-math-500Qwen3-8B + FlowBalance#42Accuracy: 93.52