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

Hugging Face Daily Papers Papers
latent-reasoning chain-of-thought llm efficiency geometric-dynamics ai-research machine-learning

Summary

A*-Thought-V2 models chain-of-thought reasoning as hidden-state trajectories, using geometric dynamics to compress non-essential steps into latent tokens, improving accuracy and efficiency in LLM reasoning.

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. We present A*-Thought-V2, a geometric dynamics of LLM guided framework that models CoT as a hidden-state trajectory and replaces hard deletion with an explicit-implicit interleaved latent architecture. After projecting question, step, and solution representations into a 3D PCA space, it measures alignment between each local transition and global question-to-solution direction. Aligned steps remain explicit text, whereas deviating steps are compressed into continuous latent tokens. Directional angles capture both local semantics and reasoning dynamics: small angles indicate direct execution and answer formation, while large angles more frequently involve checking, correction, and branch exploration; their temporal variation reveals exploration, convergence, and refinement stages. To train this architecture, we introduce stepwise embedding forcing, which pools each redundant step into a single latent embedding, and label forcing, which supervises that latent token with a soft multi-modal vocabulary distribution instead of a hard one-hot label. Experiments on Qwen3.5-9B and Qwen3.6-27B across six in-domain and out-of-domain benchmarks show that A*-Thought-V2 improves average accuracy by up to 2.6% while reducing response length by up to half, increasing Accuracy per Computation Unit by 2.29times, and reducing preprocessing and training time by 94.6% and up to 80.3%, respectively. Representation analyses suggest that latent states form a compact region distinct from textual states, while higher entropy at latent-token positions reflects broader soft targets that encourage richer step-level feature learning.
Original Article
View Cached Full Text

Cached at: 09/09/26, 12:31 PM

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

Source: https://huggingface.co/papers/2609.07821 Authors:

,

,

,

,

,

,

,

,

,

,

,

,

,

,

Abstract

A*-Thought-V2 models chain-of-thought reasoning as hidden-state trajectories to selectively retain explicit reasoning steps or compress them into continuous latent tokens, improving accuracy and efficiency.

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. We presentA*-Thought-V2, a geometric dynamics of LLM guided framework that models CoT as ahidden-state trajectoryand replaces hard deletion with anexplicit-implicit interleaved latent architecture. After projecting question, step, and solution representations into a3D PCA space, it measures alignment between each local transition and global question-to-solution direction. Aligned steps remain explicit text, whereas deviating steps are compressed into continuouslatent tokens.Directional anglescapture both local semantics and reasoning dynamics: small angles indicate direct execution and answer formation, while large angles more frequently involve checking, correction, and branch exploration; their temporal variation reveals exploration, convergence, and refinement stages. To train this architecture, we introducestepwise embedding forcing, which pools each redundant step into a single latent embedding, andlabel forcing, which supervises that latent token with asoft multi-modal vocabulary distributioninstead of a hard one-hot label. Experiments on Qwen3.5-9B and Qwen3.6-27B across six in-domain and out-of-domain benchmarks show thatA*-Thought-V2improves average accuracy by up to 2.6% while reducing response length by up to half, increasing Accuracy per Computation Unit by 2.29times, and reducing preprocessing and training time by 94.6% and up to 80.3%, respectively. Representation analyses suggest that latent states form a compact region distinct from textual states, while higher entropy at latent-token positions reflects broader soft targets that encourage richer step-level feature learning.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2609\.07821

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2609.07821 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.07821 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.07821 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles

When Do LLMs Reason? A Dynamical Systems View via Entropy Phase Transitions

arXiv cs.LG

This paper investigates when chain-of-thought reasoning is beneficial for LLMs, showing that early-stage entropy dynamics reliably indicate reasoning utility, and introduces EDRM, a lightweight, training-free framework that adaptively selects inference strategies to achieve significant token savings while maintaining or improving accuracy.

Formalizing Latent Thoughts: Four Axioms of Thought Representation in LLMs

Hugging Face Daily Papers

Introduces an axiomatic evaluation framework for latent thought representations in LLMs, revealing that current representations fail to satisfy four fundamental functional axioms (Causality, Minimality, Separability, Stability) across 23 reasoning tasks, indicating a structural gap in representation quality.

Learning to Refine Hidden States for Reliable LLM Reasoning

arXiv cs.LG

Proposes ReLAR, a reinforcement-guided latent refinement framework that iteratively updates hidden representations in LLMs before decoding, improving reasoning reliability and efficiency compared to chain-of-thought methods.

Agentic Chain-of-Thought Steering for Efficient and Controllable LLM Reasoning

Hugging Face Daily Papers

ACTS (Agentic Chain-of-Thought Steering) formulates LLM reasoning control as a Markov decision process where a controller agent adaptively steers a frozen reasoner during inference using reasoning strategies and steering phrases. The approach achieves comparable accuracy to full-thinking models with significant token savings, enabling controllable accuracy-efficiency trade-offs.