CopT: Contrastive On-Policy Thinking with Continuous Spaces for General and Agentic Reasoning

Benchmark Model Rank Results
code-generation-on-mbppCopT (Qwen3-8B)#4Accuracy: 94.55
code-generation-on-humanevalCopT (Qwen3-8B)#9Pass@1: 96.34
mathematical-reasoning-on-math-500CopT (Qwen3-8B)#9Accuracy: 97.6
arithmetic-reasoning-on-gsm8kCopT (Qwen3-8B)#10Accuracy: 96.36
mathematical-reasoning-on-aime24CopT (Qwen3-8B)#37Acc: 79.17
mathematical-reasoning-on-aime-2025CopT (Qwen3-8B)#55Accuracy: 70.42
multiple-choice-question-answering-mcqa-on-gpqa-diamondCopT (Qwen3-8B)#94Accuracy: 61.62