Learning More by Training Less: Uncertainty-Guided Progressive Contraction for Efficient Reasoning Post-Training

Benchmark Model Rank Results
arithmetic-reasoning-on-gsm8kQwen3.5-9B (rho=5%)#2Accuracy: 97.7
mathematical-reasoning-on-math-500Qwen3.5-9B (rho=30%)#4Accuracy: 98.6
mathematical-reasoning-on-aime24Qwen3.5-9B (rho=30%)#6Acc: 93.3
mathematical-reasoning-on-aime-2025Qwen3.5-9B (rho=30%)#12Accuracy: 93.3
mathematical-reasoning-on-aime-2026Qwen3.5-9B (rho=30%)#19Accuracy: 90.0