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

A*-Thought-V2 compresses chain-of-thought steps into continuous latent tokens guided by hidden-state geometry, improving average accuracy by up to 2.6% while cutting response length by up to 50% on Qwen3.5-9B and Qwen3.6-27B.

A*-Thought-V2 models chain-of-thought as a hidden-state trajectory in a projected geometric space (the arXiv report specifies a 3D PCA space) and interleaves explicit text with continuous latent tokens. Steps whose transitions deviate from the question-to-solution direction are compressed into latent tokens, while steps aligned with that direction remain explicit, replacing hard step pruning; per the arXiv report, small directional angles mark direct-execution steps and large angles mark checking and exploration steps. Training uses stepwise embedding forcing and label forcing with soft multi-modal vocabulary supervision. The method was evaluated on Qwen3.5-9B and Qwen3.6-27B across six benchmarks, described by the arXiv report as in-domain and out-of-domain. Both reports agree on the reported results: up to 2.6% average accuracy gain, responses up to 50% (up to half) shorter, 2.29x higher Accuracy per Computation Unit, 94.6% faster preprocessing, and up to 80.3% faster training.

• A*-Thought-V2 compresses chain-of-thought steps into continuous latent tokens within an explicit-implicit interleaved architecture, replacing hard step pruning (stated in the arXiv report).
• Chain-of-thought is modeled as a hidden-state trajectory in a projected geometric space; the arXiv report specifies a 3D PCA space.
• Steps whose transitions deviate from the question-to-solution direction are compressed into latent tokens; aligned steps remain explicit.
• Directional angles classify steps: small angles for direct execution, large angles for checking and exploration (stated in the arXiv report).
• Training uses stepwise embedding forcing and label forcing with soft multi-modal vocabulary supervision.
• Evaluated on Qwen3.5-9B and Qwen3.6-27B across six benchmarks (arXiv report describes them as in-domain and out-of-domain).
• Up to 2.6% average accuracy gain reported.
• Response length cut by up to 50% (up to half shorter).