ThoughtFold: Folding Reasoning Chains via Introspective Preference Learning

Benchmark Model Rank Results
arithmetic-reasoning-on-gsm8kThoughtFold (Qwen3-14B)#6Accuracy: 96.8
mathematical-reasoning-on-math-500ThoughtFold (Qwen3-8B)#12Accuracy: 97.4
mathematical-reasoning-on-aime24ThoughtFold (Qwen3-14B)#36Acc: 79.3
mathematical-reasoning-on-aime-2025ThoughtFold (Qwen3-14B)#58Accuracy: 69.7