A*-Thought-V2: Efficient Latent Reasoning via Hidden-State Trajectory Dynamics

Xiaoang Xu · hf · 2026-09-09

A-Thought-V2 models chain-of-thought reasoning as hidden-state trajectories in LLMs, using their geometric dynamics to selectively keep explicit reasoning steps or compress them into continuous latent tokens. The approach improves both accuracy and efficiency by shortening generated reasoning while preserving information-dense steps—a new entry in the latent reasoning space.

Original post →

More from Research

Research channel →