A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM
Xiaoang Xu, Siyuan Liu, Shuo Wang, Junlan Feng, Fanyu Meng, Zhu Zhang, Jixun Wang, Xiaorong Wang, et al.
cs.CL · cs.AI · cs.LG
Quick verdict
Claims geometric latent CoT compression gains up to 2.6pts accuracy at ~half length; promising, but no trained-latent baselines or significance tests.
Summary
The paper proposes A*-Thought-V2, a supervised fine-tuning framework that compresses chain-of-thought reasoning by interleaving explicit text with continuous latent tokens. After projecting question, step, and solution hidden states into a 3D PCA space, it computes the angle between each local step transition and the global question-to-solution direction: aligned steps below a 60° or 90° threshold stay textual, while deviating steps are pooled into single latent embeddings (Embedding Forcing) and supervised with step-averaged soft vocabulary distributions (Label Forcing). On Qwen3.5-9B and Qwen3.6-27B fine-tuned on OpenR1-Math-3k and evaluated on six benchmarks (Math500, AIME 2024-2026, ARC-Challenge, GPQA-Diamond), the 90° variant reaches 91.5/93.9 average accuracy versus 89.0/92.9 for same-data SFT, cuts average response length by roughly 16% versus SFT (about half versus the base backbone), and raises ACU from 0.35 to 0.80 on the 27B model; preprocessing time falls 94.6% versus A*-Thought and training time up to 80.3% versus SFT. Ablations (Table 2) show Label Forcing is critical. Main limitations: comparisons with trained latent-compression baselines (e.g., CODI, Coconut) are absent; Eq. 11's differing train/inference latent inputs contradict the stated strict consistency; the 1.0-2.5-point SFT-over gains lack significance tests; and the angle-semantics taxonomy rests on qualitative keyword inspection.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Geometry-guided step allocation (small-angle steps stay text, large-angle steps become latent) outperforms random and reversed allocation at matched compression rates (Table 2: 94.5 vs 92.8/91.4 average on AIME 2024-26, Qwen3.6-27B).
- Embedding Forcing and Label Forcing are jointly necessary: removing Label Forcing drops AIME 2024-26 average from 94.5 to 72.5 and removing both to 61.7 (Table 2).
- A*-Thought-V2 improves average accuracy by up to 2.6 points over same-data SFT while cutting response length ~16% vs SFT (about half vs the backbone) and raising ACU from 0.35 to 0.80 on Qwen3.6-27B (Table 1).
- Compression via the angle criterion is far cheaper than A* search: preprocessing falls from 5:16:22 to 0:16:57 (94.6%) and training time by up to 80.3% vs standard SFT (Table 3).
- Latent tokens occupy a compact cluster distinct from text states (Fig. 4) and show much higher predictive entropy (5.69 vs 0.26 mean on 9B), attributed to fitting broader soft multi-modal targets (Fig. 5).
Strengths
Clean same-data SFT control
Table 1 includes a 'w/ OpenR1-Math-3k' SFT row for both backbones, so the interleaved-latent effect is separable from fine-tuning per se: 90° variant 93.9 vs 92.9 average on Qwen3.6-27B with 11,785 vs 14,024 tokens.
Ablations target both novel components
Table 2 shows removing Label Forcing collapses AIME 2024-26 average from 94.5 to 72.5, removing both mechanisms to 61.7, and compares geometry-guided selection against Random Angles (92.8) and Reversed Selection (91.4) at matched ~48% compression rates.
Efficiency claims are measured, not asserted
Variance-aware ablation reporting and full hyperparameters
Weaknesses
No trained latent-compression baselines
Train/inference mismatch contradicts the consistency claim
Statistical support absent for headline gains
Angle-semantics taxonomy is post-hoc and under-validated
Minor overclaims
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This method paper (arXiv 2609.07821) introduces A*-Thought-V2, a supervised fine-tuning framework that compresses chain-of-thought (CoT) by interleaving explicit text with continuous latent tokens, replacing the hard pruning of the authors' predecessor A*-Thought (NeurIPS 2025). Mechanism: (1) project question, step, and solution hidden states into a 3D PCA space and score each step transition by its angle to the global question-to-solution direction (Eq. 1); (2) keep small-angle steps (threshold tau = 60° or 90°) as text and compress large-angle steps into one mean-pooled latent embedding per step ('Embedding Forcing', Eq. 3-5); (3) supervise each latent token with the step's averaged soft vocabulary distribution ('Label Forcing', Eq. 8-9) added to standard cross-entropy with weight lambda (Eq. 10). Headline claims: up to +2.6 points average accuracy over same-data SFT, ACU 0.35 to 0.80 on Qwen3.6-27B, 94.6% preprocessing-time reduction versus A*-Thought, and up to 80.3% training-time reduction versus SFT.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is A*-Thought-V2 in one sentence?
It is a fine-tuning method that keeps solution-aligned reasoning steps as readable text and compresses off-direction steps into single continuous latent tokens (supervised with averaged soft vocabulary targets), roughly halving response length versus the base model while slightly improving accuracy.
How much shorter and more efficient are its outputs?
On Qwen3.6-27B, the 90°-threshold variant averages 93.9% accuracy across six benchmarks with about 11,785 generated tokens per response (versus 23,390 for the base model), lifting ACU (accuracy per token cost) from 0.35 to 0.80; preprocessing is 94.6% faster than its search-based predecessor A*-Thought and training up to 80.3% faster than standard SFT.
Is there evidence the latent tokens actually preserve reasoning information?
The paper offers indirect evidence: PCA plots show latent states clustering separately from text states, latent positions show high predictive entropy (~5.7 vs ~0.3 for text) consistent with fitting broad soft targets, and an ablation where removing the label-forcing loss collapses AIME accuracy from 94.5% to 72.5%. However, no comparison against other trained latent-compression methods (e.g., CODI, Coconut) is provided.
Is code or data available, and can I retrain it?
The paper links a GitHub repository (github.com/AI9Stars/AStar-Thought) and a Hugging Face artifact (AStar-Thought-V2-OpenR1-Math-3k), with full hyperparameters in Tables 5-6; retraining the 9B/27B backbones requires roughly 8x A100 80GB GPUs for 3 epochs, so full replication is a substantial compute undertaking.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures appear only as extracted captions and axis labels; visually dependent judgments (cluster separation quality in Fig. 4, loss-curve shapes in Fig. 7, histogram detail in Fig. 10) could not be verified against rendered plots.
- Table 1's numbers were extracted as a dense flat block; column alignment was cross-checked by recomputing ACU values (they match), but individual per-benchmark cells could carry minor extraction noise.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.