acceptodds
Under review as a conference paper at ICLR 2027

From Experience to Deliberation: An Adaptive Cognitive Harness for LLM in Long-Term Traffic Simulation

Abstract

Long-term traffic simulation with large language model (LLM) agents requires stable routines and adaptation to disruptions under limited personal information. Existing agents use memory, planning, and reflection to sustain behavior, but retaining experience does not determine when acquiring information or invoking an LLM will improve an action. We propose an adaptive cognitive harness that couples persistent beliefs and habits with execution-supervised allocation of cognitive effort. Beliefs encode uncertain outcome expectations and habits encode context-dependent repetition; together with accessible evidence and constraints, they guide fast action, querying, LLM comparison, or their combination. We use paired simulator rollouts to train a lightweight predictor of task gains and resource costs relative to FAST. During deployment, the predictor remains frozen while observed outcomes revise personal state. Across five 30-day world seeds evaluated in a paired design, the hybrid policy achieves lower mean task loss than all four mandatory-LLM controls while using approximately 40–44% fewer online LLM tokens, with quality gains concentrated in the disruption period. These results establish a favorable quality–cost trade-off for selective deliberation and support an execution-grounded approach to longterm LLM simulation: cognitive effort is allocated according to decision demands rather than uniformly spent on every action. Code will be released at https://anonymous.4open.science/r/Adaptive-Cognitive-Harness-CB35/.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Traffic Simulation, Large Language Model Agents, Agent Harnesses, Belief and Habit Learning, Long-Term Behavioral Simulation

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 2 reviews, above 0% of the others.
Reviewer_wzMROpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates the allocation of cognitive computation for LLM agents in long-term traffic simulation. The authors propose an Adaptive Cognitive Harness that combines persistent belief and habit representations with a cognitive operation selection mechanism, enabling agents to dynamically select different operations based on their current states.

Critical strengths

The paper addresses a problem with practical relevance.

Critical weaknesses

1.The paper lacks key baselines to establish the independent effectiveness of the learned allocator.

  1. The improvements in task performance are marginal and lack statistical significance.

  2. The actual implementation of cost-aware allocation is not fully aligned with the methodological motivation.

  3. The paper is difficult to follow, which significantly affects its readability and clarity.

Questions for authors

Please refer to the weaknesses above.

Reviewer_BtqqOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a harness for traffic simulation that integrates belief and habit mechanisms with selective information queries and LLM-based route comparisons. A predictor determines when the operations are likely to improve upon a structured policy. The observed execution outcomes continually update each agent’s internal state. Experiments under simulated traffic disruptions show that their proposed approach achieves the lowest average task loss among baselines requiring LLM intervention and reduces LLM token cost. However, the experiments fail to isolate the contributions of personalized interaction history and the learned operation selection policy. Also, the effectiveness in real-world traffic environments remains unverified, including the sim-to-real gap and fidelity validation.

Critical strengths
  • They decouple information acquisition from LLM deliberation, which enables flexible and cost-effective cognitive allocation instead of spending expensive model calls across all scenarios.

  • Their method learns marginal gains from paired simulator counterfactuals. This grounds effort allocation in task outcomes rather than LLM-based assessment or heuristic confidence scores.

Critical weaknesses
  • The evaluation is confined to a toy synthetic setting with only 10 agents and artificially injected delays. Compared with existing works with large-scale simulations (e.g., GATSim, AgentSociety), they lack feedback on environmental congestion or calibration against real-world human mobility data. Their findings fail to demonstrate whether the mechanism works under real urban traffic dynamics.

  • The evaluation fails to isolate the individual contributions. The experiments do not establish whether the learned gate outperforms simple rule heuristics. They also fail to isolate the value of personal history over baselines with the same information.

Questions for authors
  • How does the frozen ridge predictor perform against a simple threshold rule or random gating in task loss and token savings?

  • What is the performance of only FAST (and a fixed query + FAST policy) when provided with identical report access across 30 days?

  • How does the system behave when scaled to hundreds or thousands of agents whose actions could create bottleneck congestion?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.