A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM
A*-Thought-V2 compresses chain-of-thought into latent tokens using geometric hidden-state dynamics, cutting computation while improving accuracy on Qwen models.
A*-Thought-V2 models chain-of-thought as a hidden-state trajectory and interleaves explicit text with continuous latent tokens, compressing steps whose transitions deviate from the question-to-solution direction. Trained via stepwise embedding forcing and label forcing with soft multi-modal supervision, it was evaluated on Qwen3.5-9B and Qwen3.6-27B across six benchmarks. Reported results include up to 2.6% average accuracy gain, up to 50% shorter responses, 2.29x higher Accuracy per Computation Unit, 94.6% faster preprocessing, and up to 80.3% faster training.
- Models reasoning as hidden-state trajectories in a projected geometric space
- Deviation angle from question-to-solution direction decides explicit vs latent steps
- Latent tokens trained with stepwise embedding forcing and soft label forcing
- Up to 2.6% accuracy gain with up to half the response length on Qwen models
- Preprocessing cut by 94.6% and training time by up to 80.3%
- AI models
- Qwen3.5-9BQwen3.6-27B
From the original article
1 min read at huggingface.co
Chain-of-Thought (CoT) improves the reasoning ability of Large Language Models (LLMs) but incurs substantial computation and context costs. Existing methods either lose intermediate information through hard pruning or lack a principled criterion for continuous compression.
By Xiaoang Xu, Siyuan Liu, Shuo Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.