A*-Thought-V2 Halves Reasoning Length via Latent Compression
TL;DR
- A*-Thought-V2 reports cutting response length by up to half and training time by up to 80.3%, tested on Qwen3.5-9B and Qwen3.6-27B.
- Reasoning steps aligned with the global question-to-solution direction stay as text; deviating steps are compressed into continuous latent tokens.
- Accuracy per Computation Unit rises 2.29× and preprocessing time drops 94.6% across six in-domain and out-of-domain benchmarks.
A framework called A*-Thought-V2 reports cutting large-language-model reasoning response length by up to half while trimming training time by up to 80.3%, according to a paper posted to arXiv.
The approach treats chain-of-thought as "a hidden-state trajectory" and projects each reasoning step into a 3D PCA space. Steps whose direction aligns with the global question-to-solution vector stay as explicit text; deviating steps are compressed into continuous latent tokens. In the paper's own words, the framework "replaces hard deletion with an explicit-implicit interleaved latent architecture."
Reported gains: average accuracy up by up to 2.6%, Accuracy per Computation Unit up 2.29×, and preprocessing time down 94.6%. Experiments ran on Qwen3.5-9B and Qwen3.6-27B across six in-domain and out-of-domain benchmarks.
The abstract offers a geometric read on when reasoning goes exploratory: "small angles indicate direct execution and answer formation, while large angles more frequently involve checking, correction, and branch exploration." Those deviating steps, the checking and correction, are what the method pushes into latent space.
The abstract publishes no per-benchmark numbers, no head-to-head comparisons with prior latent-CoT methods, and no wall-clock inference figures.
Originally reported by paper
Read the original article →Original headline: A*-Thought-V2 Cuts LLM Reasoning Length 50%, Training Time 80% via Geometric Trajectory Compression