A*-Thought-V2 compresses reasoning trajectories into latent tokens

Latent token compression is the missing link for scalable test-time compute, allowing models to deliberate and backtrack without the crippling latency of verbose natural language reasoning.

• –Geometric Trajectory Separation: Instead of discarding intermediate thoughts and losing context, 3D PCA mapping separates direct logical deduction from exploratory tangents and encodes the deviations into dense continuous vectors.
• –Selective Interpretability: Unlike purely implicit reasoning systems that act as impenetrable black boxes, this approach preserves core logical milestones in readable text while hiding only noisy trial-and-error.
• –Stable Continuous Optimization: Stepwise embedding forcing paired with soft vocabulary label forcing bypasses the instability of reinforcement learning over continuous tokens, reducing training time by up to 80%.
• –Proven Efficiency Gains: Demonstrating a 2.29x boost in accuracy per computation unit on open models like Qwen 3.5 and Qwen 3.6 shows that deep deliberation does not require runaway inference costs.