A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM
A*-Thought-V2 keeps some reasoning steps as text and compresses others into latent tokens based on their geometric alignment.
The paper treats chain-of-thought reasoning as a hidden-state trajectory, using angles in a 3D PCA space to decide which steps stay explicit and which become continuous latent representations. Its training setup pools redundant steps into single latent embeddings and supervises them with soft vocabulary distributions. On Qwen3.5-9B and Qwen3.6-27B, the authors report up to 2.6% average accuracy gains while cutting response length by as much as half. They also report a 2.29x gain in Accuracy per Computation Unit and large reductions in preprocessing and training time. HF Daily Papers' note
The paper treats chain-of-thought reasoning as a hidden-state trajectory, using angles in a 3D PCA space to decide which steps stay explicit and which become continuous latent representations. Its training setup pools redundant steps into single latent embeddings and supervises them with soft vocabulary distributions. On Qwen3.5-9B and Qwen3.6-27B, the authors report up to 2.6% average accuracy gains while cutting response length by as much as half. They also report a 2.29x gain in Accuracy per Computation Unit and large reductions in preprocessing and training time. HF Daily Papers' note
score 4