Geometry of Divergence: Tracking Hidden-State Trajectories for Adaptive Multi-Turn Reasoning
Jie Liang, Zhengxin Yu, Hamid Nasiri, Peter Garraghan
cs.AI · cs.CL
Quick verdict
Hidden-state curvature/slope separate agent success pre-completion and lift τ-Bench 24.1%→39.6%; promising, but threshold tuning is under-reported.
Summary
The paper formulates multi-turn LLM-agent reasoning as a hidden-state trajectory: at each turn, the hidden state at the final token of the user input is sampled from a fixed intermediate layer and summarized by two signals — temporal curvature (the cosine between consecutive turn-to-turn displacement vectors) and variance slope (the OLS slope of trajectory dispersion against prefix length). Across Math and Code sharded-dialogue tasks and τ-Bench Airline/Retail with Qwen3-14B, Qwen3-32B and Llama-3.1-8B, failed episodes show more negative curvature and lower variance slopes; separability is significant everywhere except Airline variance slope (p=0.30, p=0.057). Decomposing episodes into three-action chains (Read/Write/Respond/Transfer) shows separability concentrates in information-output chains such as Respond×3. Used as an online trigger for thinking mode, curvature thresholds (κ<−0.15/−0.20) raise mean τ-Bench reward from 0.241 to 0.396 while cutting token cost 11.2%, matching or beating a learned trigger in three of four settings. Limitations: thresholds and the κ-vs-β variant appear tuned per setting on evaluation runs, only 5 trials per task without variance reporting, small curvature effect sizes, no multiple-comparison correction, and no code release.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Temporal curvature and variance slope separate correct from incorrect episodes before completion across four tasks and three LLMs.
- Separability is action-dependent, concentrated in information-output chains such as Respond→Respond→Respond.
- Geometry-thresholded thinking triggers raise mean τ-Bench task reward from 0.241 (24.1%) to 0.396 (39.6%) while cutting token cost 11.2% (104.8k to 93.0k).
- Training-free curvature thresholds match or beat a learned trigger classifier in three of four domain-model settings (e.g., 0.397 vs 0.398 on Retail 14B at lower token cost).
Strengths
Explicit confound control via turn alignment
§4.4 identifies that incorrect episodes are systematically longer, making trajectories mechanically more curved and dispersed, and removes this by indexing one hidden state per turn before computing κ and β — methodological care many internal-state probes lack.
Training-free trigger matches a learned classifier
Table 3: κ<−0.20 ties Learned-HST on Retail/14B (0.397 vs 0.398) at lower token cost (101.0k vs 104.4k), and κ<−0.15 reaches 0.442 on Retail/32B vs the learned 0.405 — supporting the distribution-free, recalibrate-one-scalar argument.
Honest negatives profile
Fully specified, reimplementable math
Mechanistic localization of the signal
Weaknesses
Post-hoc threshold and variant selection
No uncertainty on policy comparisons
Small curvature effect sizes
Uncontrolled multiple comparisons
Under-justified probe design, no code release
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Review: Geometry of Divergence (arXiv 2608.30650)
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 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 are temporal curvature and variance slope of hidden states in this paper?
Temporal curvature is the cosine of the angle between two consecutive turn-to-turn displacement vectors of the LLM's hidden states, ranging from −1 to 1: negative values mean each update partially reverses the previous direction, positive values mean consistent accumulation. Variance slope is the regression slope of the trajectory's dispersion (trace of covariance) against the number of turns: positive means representations spread out as the conversation grows, negative means they contract.
How much does this method improve τ-Bench performance?
Averaged over four domain-model settings (Airline/Retail × Qwen3-14B/32B), thresholding hidden-state geometry to trigger thinking mode raised mean task reward from 0.241 under never-thinking to 0.396 — 24.1% to 39.6% — while reducing mean token cost by 11.2% (104.8k to 93.0k). Per-setting gains vary, and the reported average takes the better of the curvature and variance-slope variants per setting, which a careful reader should treat cautiously.
Does the method require training or an extra model?
No. The geometry-conditioned triggers are training-free: they threshold the trajectory's curvature or variance slope and toggle the LLM's built-in thinking mode on or off at each turn. The paper also evaluates a learned classifier over recent actions plus geometry, but it is not required to match performance in most settings.
Which benchmarks and models were used?
Math and Code tasks from the sharded-prompt framework of LLMs Get Lost in Multi-Turn Conversation (Laban et al. 2025) for the separability analysis, and the Airline and Retail domains of τ-Bench for both analysis and online control. Models are Qwen3-14B, Qwen3-32B, and Llama-3.1-8B for analysis; the two Qwen models for the trigger experiments.
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.
- Figure-internal distributions (boxplot shapes in Fig. 2, turn-resolved curves in Fig. 3, PCA scatter in Fig. 4c) are only partially recoverable from text extraction; judgments about visual overlap rely on the paper's quoted statistics.
- The Appendix (full experimental setup and benchmark-selection details) is referenced but not included in the provided text, so setup specifics such as trial seeding and the threshold-calibration protocol cannot be verified.
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.