Geometry of Divergence: Tracking Hidden-State Trajectories for Adaptive Multi-Turn Reasoning

Public
Authors:Jie Liang, Zhengxin Yu, Hamid Nasiri, Peter Garraghan
More
Copy repository link

AI research summary

LLM agents need to sustain goal-consistent reasoning across long multi-turn interactions under strict resource constraints. However, as the multi-turn context accumulates, it can destabilize the underlying LLM's internal representation of task-relevant information from earlier turns, blurring the boundary between constructive reasoning and representation drift. We formulate multi-turn reasoning as a hidden-state trajectory of the underlying LLM that is characterized via two complementary signals: temporal curvature that captures the directional consistency of turn-to-turn updates, and variance slope which measures the expansion or contraction of the exploration space. Across four tasks and three underlying LLMs, we observed that these geometric signals distinguish between correct and incorrect episodes prior to completion. We further decompose each episode into three-action chains formed from four actions (Read, Write, Respond, Transfer) and show that separability is action-dependent, with different signals distinguishing various chain patterns. Our experiments demonstrate that trajectory geometry can identify critical turns in the reasoning process, increasing task success rates on tau-Bench from 24.1% to 39.6% while reducing token cost by 11.2%.

Generated by CiteArk.

What this means

Autonomous LLM agents frequently fail during multi-turn interactions because accumulating interaction history destabilizes their internal representations, causing context saturation, representation drift, and unneeded reasoning overhead. This paper proposes tracking the trajectory of intermediate hidden states across conversational turns using two geometric metrics: temporal curvature (measuring directional consistency of hidden-state displacements) and variance slope (measuring expansion or contraction of the exploration space). Instead of relying on static context engineering or always-on chain-of-thought deliberation, the authors demonstrate that monitoring these latent geometric signals can trigger deliberative reasoning dynamically only when failure or drift is imminent. For practitioners and researchers developing conversational agents in tool-grounded domains like retail and airline customer support, this method offers a potential route to improve task success rates while lowering overall inference token costs. A primary operational limitation is that extracting hidden states requires direct white-box access to the model's internal activations at specific intermediate layers, making the approach inaccessible for closed-source proprietary API-only models.

Reproduction progress
RunMatchRepeat

Not reproduced yet

0/14

No verifiable claim has successful reproduction evidence yet.

Awaiting author or institution signature.

Research claims

The main empirical claims extracted in this research plan, with their plan coverage and live evidence state. Click a claim to expand it inline.

Failed episodes exhibit systematically more negative temporal curvature kappa across all six task-LLM pairings, with correct minus incorrect mean differences of: Math (Llama-3.1-8B) +0.047, Code (Qwen3-14B) +0.148, Retail (Qwen3-14B) +0.030, Airline (Qwen3-14B) +0.049, Retail (Qwen3-32B) +0.035, and Airline (Qwen3-32B) +0.088.Experiment blocked; see the specific reasonReported 0.047 scalar

Reported

0.047 scalar

Observed

The offline episode traces, prompt sharding scripts for Math and Code from Lost in Conversation, and exact hidden state trajectory extraction codebase are not provided in the paper snapshot or any public repository.

curvature_difference_correct_minus_incorrect · Not assessed

Reported: 0.047 scalar

curvature_difference_correct_minus_incorrect · Not assessed

Reported: 0.148 scalar

curvature_difference_correct_minus_incorrect · Not assessed

Reported: 0.03 scalar

curvature_difference_correct_minus_incorrect · Not assessed

Reported: 0.049 scalar

curvature_difference_correct_minus_incorrect · Not assessed

Reported: 0.035 scalar

curvature_difference_correct_minus_incorrect · Not assessed

Reported: 0.088 scalar

The learned classifier predicting the agent's next action achieves F1 scores of 0.826 for Read, 0.711 for Respond, 0.700 for Write, and 0.531 for Transfer.Experiment blocked; see the specific reasonReported 0.826 fraction

Reported

0.826 fraction

Observed

The paper does not provide the classifier model, training code, loss function, evaluation split, or feature representation.

f1_score · Not assessed

Reported: 0.826 fraction

f1_score · Not assessed

Reported: 0.711 fraction

f1_score · Not assessed

Reported: 0.7 fraction

f1_score · Not assessed

Reported: 0.531 fraction

Action-conditioned trigger policies exhibit domain-dependent performance: on Retail with Qwen3-14B, triggering on Read (0.393) and Write (0.381) outperforms Respond (0.290), whereas on Airline with Qwen3-14B, the ordering inverts, with Respond achieving 0.372 while Read achieves 0.352 and Write achieves 0.340.Experiment blocked; see the specific reasonReported 0.393 fraction

Reported

0.393 fraction

Observed

The learned action classifier (predicting next action a_hat_t) is not released, nor are the training data, features, architecture, or tau-Bench integration harness.

task_reward · Not assessed

Reported: 0.393 fraction

task_reward · Not assessed

Reported: 0.381 fraction

task_reward · Not assessed

Reported: 0.29 fraction

task_reward · Not assessed

Reported: 0.372 fraction

task_reward · Not assessed

Reported: 0.352 fraction

task_reward · Not assessed

Reported: 0.34 fraction

In the analysis of top-20 three-action chains, the information-acquisition chain Read->Read->Read (ID 1) accounts for 19.6%/22.2% (Correct/Incorrect, Rank #1) in Retail and 17.1%/14.5% (Rank #2) in Airline, while the repetitive output chain Resp->Resp->Resp (ID 2) accounts for 3.6%/8.2% (Rank #5) in Retail and 9.1%/23.6% (Rank #1) in Airline, showing strong error concentration in ungrounded conversational loops.Experiment blocked; see the specific reasonReported 19.6 pp

Reported

19.6 pp

Observed

Complete sliding window logs across episodes and domain trajectories are not available.

chain_window_share · Not assessed

Reported: 19.6 pp

chain_window_share · Not assessed

Reported: 22.2 pp

chain_window_share · Not assessed

Reported: 17.1 pp

chain_window_share · Not assessed

Reported: 14.5 pp

chain_window_share · Not assessed

Reported: 3.6 pp

chain_window_share · Not assessed

Reported: 8.2 pp

chain_window_share · Not assessed

Reported: 9.1 pp

chain_window_share · Not assessed

Reported: 23.6 pp

Correct episodes exhibit higher variance slopes beta_t than incorrect episodes across all six task-LLM pairings. Separability is statistically significant (p < .001) for Math (Llama-3.1-8B), Code (Qwen3-14B), Retail (Qwen3-14B), and Retail (Qwen3-32B), but not statistically significant for Airline (Qwen3-14B, p=0.30; Qwen3-32B, p=0.057).Experiment blocked; see the specific reasonReported 0.3 scalar

Reported

0.3 scalar

Observed

Raw episode traces, prompt sharding scripts, and hidden state trajectory extraction codebase are unreleased.

mann_whitney_u_p_value · Not assessed

Reported: 0.3 scalar

mann_whitney_u_p_value · Not assessed

Reported: 0.057 scalar

Across four domain-model settings on tau-Bench, geometry-conditioned adaptive reasoning triggers raise average task reward from 0.241 (24.1%) under Never-thinking to 0.396 (39.6%) while reducing mean token cost from 104.8k to 93.0k, an 11.2% reduction.Experiment blocked; see the specific reasonReported 0.241 fraction

Reported

0.241 fraction

Observed

The paper does not provide an official repository or executable code release (repository-identity.json indicates available: false). Furthermore, key protocol components including prompt templates, user simulator conversation driver scripts, probe extraction and logging pipelines, the exact classifier training split for Learned-HST, and detailed per-task configurations are deferred to an unreleased Appendix.

average_task_reward · Not assessed

Reported: 0.241 fraction

average_task_reward · Not assessed

Reported: 0.396 fraction

mean_token_cost_thousands · Not assessed

Reported: 104.8 thousands_tokens

mean_token_cost_thousands · Not assessed

Reported: 93 thousands_tokens

token_cost_reduction · Not assessed

Reported: 11.2 pp

Reproduction and technical details0
Implementation path
Official implementation0

The experiment runs a repository and commit explicitly declared by the paper and verified by CiteArk.

CiteArk reconstruction0

No author code is used. CiteArk independently rebuilds the paper-grounded protocol and records every generated file and assumption.

Information insufficient14

A falsifiable execution still needs a missing method detail, input, split, metric, or evaluation rule.

Signed research planVerifiedThe Research Plan CAP fixes paper sources, complete Claim coverage, planned Experiments, and source-declared values.Download research plan
Generated byCiteArk Official
Modelgoogle/gemini-3.8-flash
Completed
Artifact digest3b707a824b
Attestation statusTrusted signature verified
Output
View independent verification JSON