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.Whole claim: Experiment blocked; see the specific reason0/6 measurements supported · 0 assessed6measurementsExplore results

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.

Measurement results

0/6 assessed

0 supported0 disputed or conflicting6 not assessed

The complete result group stays with its claim. Assessed rows come first; their judgments do not replace the whole-claim conclusion. Expand a row for conditions, sources, and evidence history.

Each row is a measurement with its reported value, latest available observation, and cumulative evidence state.
Metric and conditions	Reported	Latest observation	Evidence judgment
task: Math · model: Llama-3.1-8B	0.047 scalar	—	Not assessed
task: Code · model: Qwen3-14B	0.148 scalar	—	Not assessed
task: Retail · model: Qwen3-14B	0.03 scalar	—	Not assessed
task: Airline · model: Qwen3-14B	0.049 scalar	—	Not assessed
task: Retail · model: Qwen3-32B	0.035 scalar	—	Not assessed
task: Airline · model: Qwen3-32B	0.088 scalar	—	Not assessed

Open claim details

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.Whole claim: Experiment blocked; see the specific reason0/4 measurements supported · 0 assessed4measurementsExplore results

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

Measurement results

0/4 assessed

0 supported0 disputed or conflicting4 not assessed

The complete result group stays with its claim. Assessed rows come first; their judgments do not replace the whole-claim conclusion. Expand a row for conditions, sources, and evidence history.

Each row is a measurement with its reported value, latest available observation, and cumulative evidence state.
Metric and conditions	Reported	Latest observation	Evidence judgment
action: Read	0.826 fraction	—	Not assessed
action: Respond	0.711 fraction	—	Not assessed
action: Write	0.7 fraction	—	Not assessed
action: Transfer	0.531 fraction	—	Not assessed

Open claim details

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.Whole claim: Experiment blocked; see the specific reason0/6 measurements supported · 0 assessed6measurementsExplore results

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.

Measurement results

0/6 assessed

0 supported0 disputed or conflicting6 not assessed

The complete result group stays with its claim. Assessed rows come first; their judgments do not replace the whole-claim conclusion. Expand a row for conditions, sources, and evidence history.

Each row is a measurement with its reported value, latest available observation, and cumulative evidence state.
Metric and conditions	Reported	Latest observation	Evidence judgment
model: Qwen3-14B · domain: Retail · trigger_action: Read	0.393 fraction	—	Not assessed
model: Qwen3-14B · domain: Retail · trigger_action: Write	0.381 fraction	—	Not assessed
model: Qwen3-14B · domain: Retail · trigger_action: Respond	0.29 fraction	—	Not assessed
model: Qwen3-14B · domain: Airline · trigger_action: Respond	0.372 fraction	—	Not assessed
model: Qwen3-14B · domain: Airline · trigger_action: Read	0.352 fraction	—	Not assessed
model: Qwen3-14B · domain: Airline · trigger_action: Write	0.34 fraction	—	Not assessed

Open claim details

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.Whole claim: Experiment blocked; see the specific reason0/8 measurements supported · 0 assessed8measurementsExplore results

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

Measurement results

0/8 assessed

0 supported0 disputed or conflicting8 not assessed

The complete result group stays with its claim. Assessed rows come first; their judgments do not replace the whole-claim conclusion. Expand a row for conditions, sources, and evidence history.

Each row is a measurement with its reported value, latest available observation, and cumulative evidence state.
Metric and conditions	Reported	Latest observation	Evidence judgment
group: correct · domain: Retail · pattern: Read->Read->Read · chain_id: 1	19.6 percentage_points	—	Not assessed
group: incorrect · domain: Retail · pattern: Read->Read->Read · chain_id: 1	22.2 percentage_points	—	Not assessed
group: correct · domain: Airline · pattern: Read->Read->Read · chain_id: 1	17.1 percentage_points	—	Not assessed
group: incorrect · domain: Airline · pattern: Read->Read->Read · chain_id: 1	14.5 percentage_points	—	Not assessed
group: correct · domain: Retail · pattern: Resp->Resp->Resp · chain_id: 2	3.6 percentage_points	—	Not assessed
group: incorrect · domain: Retail · pattern: Resp->Resp->Resp · chain_id: 2	8.2 percentage_points	—	Not assessed
group: correct · domain: Airline · pattern: Resp->Resp->Resp · chain_id: 2	9.1 percentage_points	—	Not assessed
group: incorrect · domain: Airline · pattern: Resp->Resp->Resp · chain_id: 2	23.6 percentage_points	—	Not assessed

Open claim details

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).Whole claim: Experiment blocked; see the specific reason0/2 measurements supported · 0 assessed2measurementsExplore results

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

Measurement results

0/2 assessed

0 supported0 disputed or conflicting2 not assessed

The complete result group stays with its claim. Assessed rows come first; their judgments do not replace the whole-claim conclusion. Expand a row for conditions, sources, and evidence history.

Each row is a measurement with its reported value, latest available observation, and cumulative evidence state.
Metric and conditions	Reported	Latest observation	Evidence judgment
task: Airline · model: Qwen3-14B	0.3 scalar	—	Not assessed
task: Airline · model: Qwen3-32B	0.057 scalar	—	Not assessed

Open claim details

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.Whole claim: Experiment blocked; see the specific reason0/5 measurements supported · 0 assessed5measurementsExplore results

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.

Measurement results

0/5 assessed

0 supported0 disputed or conflicting5 not assessed

The complete result group stays with its claim. Assessed rows come first; their judgments do not replace the whole-claim conclusion. Expand a row for conditions, sources, and evidence history.

Each row is a measurement with its reported value, latest available observation, and cumulative evidence state.
Metric and conditions	Reported	Latest observation	Evidence judgment
policy: never-think · benchmark: tau-bench	0.241 fraction	—	Not assessed
policy: geometry-conditioned · benchmark: tau-bench	0.396 fraction	—	Not assessed
policy: never-think · benchmark: tau-bench	104.8 thousands_tokens	—	Not assessed
policy: geometry-conditioned · benchmark: tau-bench	93 thousands_tokens	—	Not assessed
benchmark: tau-bench	11.2 percentage_points	—	Not assessed

Open claim details

On tau-Bench Airline with Qwen3-32B, trigger policies yield task rewards of 0.136 (Never-think), 0.280 (Warm-up), 0.364 (Always-think), 0.328 (Variance slope beta_t > tau_beta), 0.352 (Temporal curvature kappa_t < -0.10), and 0.364 (Learned-HST), with mean token costs in thousands of 98.2, 92.6, 100.6, 86.3, 83.0, and 87.5 respectively.Whole claim: Experiment blocked; see the specific reason0/18 measurements supported · 0 assessed18measurementsExplore results

Official execution artifacts, dialogue simulators, user simulator policy prompts, and trajectory monitoring hooks are not released.

Measurement results

0/18 assessed

0 supported0 disputed or conflicting18 not assessed

The complete result group stays with its claim. Assessed rows come first; their judgments do not replace the whole-claim conclusion. Expand a row for conditions, sources, and evidence history.

Each row is a measurement with its reported value, latest available observation, and cumulative evidence state.
Metric and conditions	Reported	Latest observation	Evidence judgment
model: Qwen3-32B · domain: Airline · policy: Never-think	0.136 fraction	—	Not assessed
model: Qwen3-32B · domain: Airline · policy: Warm-up	0.28 fraction	—	Not assessed
model: Qwen3-32B · domain: Airline · policy: Always-think	0.364 fraction	—	Not assessed
model: Qwen3-32B · domain: Airline · policy: Variance-slope	0.328 fraction	—	Not assessed
model: Qwen3-32B · domain: Airline · policy: Temporal-curvature · threshold: -0.1	0.352 fraction	—	Not assessed
model: Qwen3-32B · domain: Airline · policy: Learned-HST	0.364 fraction	—	Not assessed
model: Qwen3-32B · domain: Airline · policy: Never-think	98.2 thousands_tokens	—	Not assessed
model: Qwen3-32B · domain: Airline · policy: Warm-up	92.6 thousands_tokens	—	Not assessed
model: Qwen3-32B · domain: Airline · policy: Always-think	100.6 thousands_tokens	—	Not assessed
model: Qwen3-32B · domain: Airline · policy: Variance-slope	86.3 thousands_tokens	—	Not assessed
model: Qwen3-32B · domain: Airline · policy: Temporal-curvature · threshold: -0.1	83 thousands_tokens	—	Not assessed
model: Qwen3-32B · domain: Airline · policy: Learned-HST	87.5 thousands_tokens	—	Not assessed

Open claim details

On tau-Bench Airline with Qwen3-14B, trigger policies yield task rewards of 0.196 (Never-think), 0.348 (Warm-up), 0.356 (Always-think), 0.392 (Variance slope beta_t > tau_beta), 0.384 (Temporal curvature kappa_t < -0.15), and 0.420 (Learned-HST), with mean token costs in thousands of 82.5, 87.1, 83.5, 83.3, 78.9, and 85.0 respectively.Whole claim: Experiment blocked; see the specific reason0/18 measurements supported · 0 assessed18measurementsExplore results

Official execution artifacts, dialogue simulators, user simulator policy prompts, and trajectory monitoring hooks are not released.

Measurement results

0/18 assessed

0 supported0 disputed or conflicting18 not assessed

The complete result group stays with its claim. Assessed rows come first; their judgments do not replace the whole-claim conclusion. Expand a row for conditions, sources, and evidence history.

Each row is a measurement with its reported value, latest available observation, and cumulative evidence state.
Metric and conditions	Reported	Latest observation	Evidence judgment
model: Qwen3-14B · domain: Airline · policy: Never-think	0.196 fraction	—	Not assessed
model: Qwen3-14B · domain: Airline · policy: Warm-up	0.348 fraction	—	Not assessed
model: Qwen3-14B · domain: Airline · policy: Always-think	0.356 fraction	—	Not assessed
model: Qwen3-14B · domain: Airline · policy: Variance-slope	0.392 fraction	—	Not assessed
model: Qwen3-14B · domain: Airline · policy: Temporal-curvature · threshold: -0.15	0.384 fraction	—	Not assessed
model: Qwen3-14B · domain: Airline · policy: Learned-HST	0.42 fraction	—	Not assessed
model: Qwen3-14B · domain: Airline · policy: Never-think	82.5 thousands_tokens	—	Not assessed
model: Qwen3-14B · domain: Airline · policy: Warm-up	87.1 thousands_tokens	—	Not assessed
model: Qwen3-14B · domain: Airline · policy: Always-think	83.5 thousands_tokens	—	Not assessed
model: Qwen3-14B · domain: Airline · policy: Variance-slope	83.3 thousands_tokens	—	Not assessed
model: Qwen3-14B · domain: Airline · policy: Temporal-curvature · threshold: -0.15	78.9 thousands_tokens	—	Not assessed
model: Qwen3-14B · domain: Airline · policy: Learned-HST	85 thousands_tokens	—	Not assessed

Open claim details

On tau-Bench Retail with Qwen3-32B, trigger policies yield task rewards of 0.347 (Never-think), 0.405 (Warm-up), 0.384 (Always-think), 0.412 (Variance slope beta_t > tau_beta), 0.442 (Temporal curvature kappa_t < -0.15), and 0.405 (Learned-HST), with mean token costs in thousands of 125.7, 107.1, 119.3, 105.1, 104.6, and 108.5 respectively.Whole claim: Experiment blocked; see the specific reason0/18 measurements supported · 0 assessed18measurementsExplore results

Official execution artifacts, dialogue simulators, user simulator policy prompts, and trajectory monitoring hooks are not released.

Measurement results

0/18 assessed

0 supported0 disputed or conflicting18 not assessed

The complete result group stays with its claim. Assessed rows come first; their judgments do not replace the whole-claim conclusion. Expand a row for conditions, sources, and evidence history.

Each row is a measurement with its reported value, latest available observation, and cumulative evidence state.
Metric and conditions	Reported	Latest observation	Evidence judgment
model: Qwen3-32B · domain: Retail · policy: Never-think	0.347 fraction	—	Not assessed
model: Qwen3-32B · domain: Retail · policy: Warm-up	0.405 fraction	—	Not assessed
model: Qwen3-32B · domain: Retail · policy: Always-think	0.384 fraction	—	Not assessed
model: Qwen3-32B · domain: Retail · policy: Variance-slope	0.412 fraction	—	Not assessed
model: Qwen3-32B · domain: Retail · policy: Temporal-curvature · threshold: -0.15	0.442 fraction	—	Not assessed
model: Qwen3-32B · domain: Retail · policy: Learned-HST	0.405 fraction	—	Not assessed
model: Qwen3-32B · domain: Retail · policy: Never-think	125.7 thousands_tokens	—	Not assessed
model: Qwen3-32B · domain: Retail · policy: Warm-up	107.1 thousands_tokens	—	Not assessed
model: Qwen3-32B · domain: Retail · policy: Always-think	119.3 thousands_tokens	—	Not assessed
model: Qwen3-32B · domain: Retail · policy: Variance-slope	105.1 thousands_tokens	—	Not assessed
model: Qwen3-32B · domain: Retail · policy: Temporal-curvature · threshold: -0.15	104.6 thousands_tokens	—	Not assessed
model: Qwen3-32B · domain: Retail · policy: Learned-HST	108.5 thousands_tokens	—	Not assessed

Open claim details

Sweeping the temporal curvature threshold tau_kappa on Retail with Qwen3-14B yields task rewards of 0.388 at tau_kappa = -0.15, 0.397 at tau_kappa = -0.20, and 0.363 at tau_kappa = -0.25, demonstrating unimodal behavior.Whole claim: Experiment blocked; see the specific reason0/3 measurements supported · 0 assessed3measurementsExplore results

Evaluation scripts and environment setup for sweeping tau_kappa on tau-Bench Retail are not available.

Measurement results

0/3 assessed

0 supported0 disputed or conflicting3 not assessed

The complete result group stays with its claim. Assessed rows come first; their judgments do not replace the whole-claim conclusion. Expand a row for conditions, sources, and evidence history.

Each row is a measurement with its reported value, latest available observation, and cumulative evidence state.
Metric and conditions	Reported	Latest observation	Evidence judgment
model: Qwen3-14B · domain: Retail · threshold: -0.15	0.388 fraction	—	Not assessed
model: Qwen3-14B · domain: Retail · threshold: -0.2	0.397 fraction	—	Not assessed
model: Qwen3-14B · domain: Retail · threshold: -0.25	0.363 fraction	—	Not assessed

Open claim details

In three-action chains, the proportions of chains containing each action differ between correct and incorrect episodes. In Retail, Write occurs in 49.3% vs 43.4% (Delta = -5.9%), Read in 85.8% vs 82.5% (Delta = -3.3%), Transfer in 4.2% vs 4.4% (Delta = +0.2%), and Respond in 57.4% vs 50.0% (Delta = -7.4%). In Airline, Write occurs in 15.3% vs 38.9% (Delta = +23.7%), Read in 86.1% vs 64.4% (Delta = -21.7%), Transfer in 29.2% vs 7.5% (Delta = -21.7%), and Respond in 58.3% vs 61.7% (Delta = +3.4%).Whole claim: Experiment blocked; see the specific reason0/16 measurements supported · 0 assessed16measurementsExplore results

The complete execution traces from which three-action sliding windows were extracted are not published, and action labeling mappings for tool calls are unspecified.

Measurement results

0/16 assessed

0 supported0 disputed or conflicting16 not assessed

The complete result group stays with its claim. Assessed rows come first; their judgments do not replace the whole-claim conclusion. Expand a row for conditions, sources, and evidence history.

Each row is a measurement with its reported value, latest available observation, and cumulative evidence state.
Metric and conditions	Reported	Latest observation	Evidence judgment
group: correct · action: Write · domain: Retail	49.3 percentage_points	—	Not assessed
group: incorrect · action: Write · domain: Retail	43.4 percentage_points	—	Not assessed
group: correct · action: Read · domain: Retail	85.8 percentage_points	—	Not assessed
group: incorrect · action: Read · domain: Retail	82.5 percentage_points	—	Not assessed
group: correct · action: Transfer · domain: Retail	4.2 percentage_points	—	Not assessed
group: incorrect · action: Transfer · domain: Retail	4.4 percentage_points	—	Not assessed
group: correct · action: Respond · domain: Retail	57.4 percentage_points	—	Not assessed
group: incorrect · action: Respond · domain: Retail	50 percentage_points	—	Not assessed
group: correct · action: Write · domain: Airline	15.3 percentage_points	—	Not assessed
group: incorrect · action: Write · domain: Airline	38.9 percentage_points	—	Not assessed
group: correct · action: Read · domain: Airline	86.1 percentage_points	—	Not assessed
group: incorrect · action: Read · domain: Airline	64.4 percentage_points	—	Not assessed

Open claim details

On the Code task with Qwen3-14B, the correct group exhibits a positive median temporal curvature of +0.127 and a mean of +0.096, indicating that hidden-state updates on structured verifiable tasks tend toward same-direction accumulation rather than directional reversal.Whole claim: Experiment blocked; see the specific reason0/2 measurements supported · 0 assessed2measurementsExplore results

The Code task generation, interaction logs, and hidden-state extraction code are not available.

Measurement results

0/2 assessed

0 supported0 disputed or conflicting2 not assessed

The complete result group stays with its claim. Assessed rows come first; their judgments do not replace the whole-claim conclusion. Expand a row for conditions, sources, and evidence history.

Each row is a measurement with its reported value, latest available observation, and cumulative evidence state.
Metric and conditions	Reported	Latest observation	Evidence judgment
task: Code · group: correct · model: Qwen3-14B	0.127 scalar	—	Not assessed
task: Code · group: correct · model: Qwen3-14B	0.096 scalar	—	Not assessed

Open claim details

In the turn-resolved hidden-state trajectory analysis on Math tasks (Figure 3), the sample sizes comprise n=259 correct episodes and n=227 incorrect episodes.Whole claim: Experiment blocked; see the specific reason0/2 measurements supported · 0 assessed2measurementsExplore results

The exact Math problem subset, seeds, sharding prompt splits, and model generation outputs from Lost in Conversation are not included in the paper snapshot.

Measurement results

0/2 assessed

0 supported0 disputed or conflicting2 not assessed

The complete result group stays with its claim. Assessed rows come first; their judgments do not replace the whole-claim conclusion. Expand a row for conditions, sources, and evidence history.

Each row is a measurement with its reported value, latest available observation, and cumulative evidence state.
Metric and conditions	Reported	Latest observation	Evidence judgment
task: Math · group: correct	259 count	—	Not assessed
task: Math · group: incorrect	227 count	—	Not assessed

Open claim details

On tau-Bench Retail with Qwen3-14B, trigger policies yield task rewards of 0.286 (Never-think), 0.361 (Warm-up), 0.307 (Always-think), 0.393 (Variance slope beta_t > tau_beta), 0.397 (Temporal curvature kappa_t < -0.20), and 0.398 (Learned-HST), with mean token costs in thousands of 112.6, 111.1, 113.1, 104.3, 101.0, and 104.4 respectively.Whole claim: Experiment blocked; see the specific reason0/18 measurements supported · 0 assessed18measurementsExplore results

Execution requires official code and prompt/dialogue harnesses that are not published with the paper. Missing items include the tau-Bench agent loop modifications, the thinking mode activation protocol for Qwen3-14B, the layer 22 extraction hook, and the specific Learned-HST classifier model and weights.

Measurement results

0/18 assessed

0 supported0 disputed or conflicting18 not assessed

The complete result group stays with its claim. Assessed rows come first; their judgments do not replace the whole-claim conclusion. Expand a row for conditions, sources, and evidence history.

Each row is a measurement with its reported value, latest available observation, and cumulative evidence state.
Metric and conditions	Reported	Latest observation	Evidence judgment
model: Qwen3-14B · domain: Retail · policy: Never-think	0.286 fraction	—	Not assessed
model: Qwen3-14B · domain: Retail · policy: Warm-up	0.361 fraction	—	Not assessed
model: Qwen3-14B · domain: Retail · policy: Always-think	0.307 fraction	—	Not assessed
model: Qwen3-14B · domain: Retail · policy: Variance-slope	0.393 fraction	—	Not assessed
model: Qwen3-14B · domain: Retail · policy: Temporal-curvature · threshold: -0.2	0.397 fraction	—	Not assessed
model: Qwen3-14B · domain: Retail · policy: Learned-HST	0.398 fraction	—	Not assessed
model: Qwen3-14B · domain: Retail · policy: Never-think	112.6 thousands_tokens	—	Not assessed
model: Qwen3-14B · domain: Retail · policy: Warm-up	111.1 thousands_tokens	—	Not assessed
model: Qwen3-14B · domain: Retail · policy: Always-think	113.1 thousands_tokens	—	Not assessed
model: Qwen3-14B · domain: Retail · policy: Variance-slope	104.3 thousands_tokens	—	Not assessed
model: Qwen3-14B · domain: Retail · policy: Temporal-curvature · threshold: -0.2	101 thousands_tokens	—	Not assessed
model: Qwen3-14B · domain: Retail · policy: Learned-HST	104.4 thousands_tokens	—	Not assessed

Open claim details

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

Research claims

Reproduction results

No reproduction results yet

Key metricsPaper-reported → reproduced · click a row to expand

Measurement results

Measurement results

Measurement results

Measurement results

Measurement results

Measurement results

Measurement results

Measurement results

Measurement results

Measurement results

Measurement results

Measurement results

Measurement results

Measurement results