Reproduce “A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM” on OpenResearch

About this paper

Xiaoang Xu, Siyuan Liu, Shuo Wang, Junlan Feng, Fanyu Meng, Zhu Zhang, Jixun Wang, Xiaorong Wang and 8 more

A*-Thought-V2 is a framework for efficient Chain-of-Thought (CoT) reasoning in large language models that leverages geometric dynamics of hidden states to dynamically compress explicit reasoning steps into continuous latent representations. This approach reduced average response length by 16.0% and improved Accuracy per Computation Unit (ACU) by up to 2.29x on the Qwen3.6-27B model across various benchmarks, while maintaining or enhancing reasoning accuracy.

Key insights

• The geometric dynamics of hidden states, specifically directional angles relative to the global solution path, provide a semantically grounded criterion for identifying and compressing less direct reasoning steps.
• Encoding 'redundant' reasoning segments into dense, continuous latent tokens (rather than discarding them) preserves more information while significantly reducing sequence length and computational load.
• Both stepwise embedding forcing and, critically, label forcing with soft targets are essential for effectively training LLMs to learn and utilize rich, dense latent representations for compressed reasoning.

Summary by alphaXiv.