A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM

Benchmark Model Rank Results
mathematical-reasoning-on-aime24A*-Thought-V2 (Qwen3.6-27B, tau=90)#1Acc: 96.7
mathematical-reasoning-on-aime-2025A*-Thought-V2 (Qwen3.6-27B, tau=60)#9Accuracy: 95.0
mathematical-reasoning-on-aime-2026A*-Thought-V2 (Qwen3.6-27B, tau=90)#13Accuracy: 93.3
mathematical-reasoning-on-math-500A*-Thought-V2 (Qwen3.5-9B, tau=60)#16Accuracy: 97.0
multiple-choice-question-answering-mcqa-on-gpqa-diamondA*-Thought-V2 (Qwen3.6-27B, tau=90)#31Accuracy: 86.0