Bridging Local Observation and Global Simulation in Closed-Loop Traffic Modeling

Ziyan Wang1,2* · Tan Xiang2* · Peng Chen2 · Xintao Yan1†

1The University of Hong Kong · 2Beihang University · *Equal contribution · †Corresponding author

CRAFT, a Contextual pReference Alignment Framework for Traffic Simulation, bridges this local-to-global mismatch via self-supervised failure discovery and preference-guided test-time alignment.

Read the paper arXiv Code View rollouts ↓
31.2% lower scene-level collision rate
for CAT-K
33.2% lower traffic-violation rate
for SMART
48.7% of abnormal behaviors warned
at least 1.0 s early
23.79% lower abnormal-agent rate
at the 9 s horizon

Overview

Can policies learned from locally observed driving logs remain behaviorally rational when executed in globally contextualized closed-loop simulation?

Autoregressive traffic simulators learn from ego-centric driving logs, where surrounding context can be incomplete due to perception limits and occlusions. When these models are rolled out in globally observable simulation, incomplete context-action associations can lead to abnormal stops, unsafe interactions, and rule violations.

Local-to-global context mismatch and CRAFT alignment
Incomplete ego-centric logs can induce flawed context-action mappings. CRAFT mitigates this mismatch by learning complete-context preferences and guiding simulation toward more rational behaviors.

Method

CRAFT turns simulator-induced failures into preference guidance.

CRAFT treats the simulator itself as a globally observable sandbox. Starting from logged initial states, the base simulator generates diverse what-if rollouts that self-expose failure modes induced by incomplete observational context. These failures are grounded with human-aligned driving priors and converted into preference supervision for a Contextual Preference Evaluator (CPE), which assesses the rationality of generated behaviors under complete scene context.

Training

Construct grouped complete-context what-if rollouts.

The base simulator performs stochastic rollouts with diverse random seeds from the same logged initial scene. The resulting preference dataset is organized as grouped alternative futures, supporting token-level supervision and inter-rollout preference learning.

CRAFT training pipeline

Inference

Reweight autoregressive decoding before action execution.

At each simulation step, CPE evaluates next-step candidates through one-step lookahead scenes and calibrates the simulator's next-token distribution before action execution.

CRAFT inference pipeline

Qualitative Rollouts

CRAFT guides closed-loop rollouts toward globally coherent traffic behavior.

Each scenario compares the baseline simulator with CRAFT under the same initial scene. CRAFT uses complete-context preference guidance to suppress locally plausible but globally inconsistent actions.

Baseline
CRAFT

Traffic-signal compliance

The baseline selects an action that violates the traffic signal, while CRAFT suppresses the violating candidate and maintains rule-compliant behavior.

Baseline
CRAFT

Abnormal-stop mitigation

The baseline produces an unexplained stop under the complete simulated context, whereas CRAFT guides the rollout toward smoother forward motion and preserves traffic flow.

Baseline
CRAFT

Safer multi-agent interaction

The baseline rollout leads to an unsafe interaction pattern, while CRAFT recalibrates candidate actions using contextual preference scores and produces a more coherent multi-agent behavior.

Baseline
CRAFT

Self-recovery under initial perturbations

The baseline struggles to recover from the initial perturbation, while CRAFT guides the vehicle back toward normal driving behavior under the complete simulated context.

Open-loop Eval.
Open-loop Eval.

Open-loop contextual preference evaluation

CPE estimates agent-time preference scores for generated behaviors under the complete simulated scene. In this visualization, green indicates a high preference score close to 1, red indicates a low preference score close to 0, and yellow indicates an intermediate score around 0.5, with colors gradually transitioning from green to red.

Main results

CRAFT improves closed-loop behavioral rationality.

CRAFT improves closed-loop behavioral rationality while maintaining competitive distributional realism. CPE functions as a plug-in alignment module for different simulators, amortizing human-aligned criteria into dense contextual preference scores and enabling more stable probability calibration during decoding.

Base Model Strategy JSD (× 10−2) ↓ Collision (%) ↓ Offroad (%) ↓ Traffic (%) ↓ P-Sc.
Spd. Ang. Dist. P-Ag. P-Sc. P-Ag. P-Sc.
Log Log replay 0.00 0.00 0.00 0.56 5.20 1.59 14.40 20.70
GUMP Top-K 4.78 5.41 11.36 4.06 36.30 3.80 26.90 32.70
SMART Top-K 1.07 3.23 0.56 3.13 15.10 2.52 19.10 38.80
CAT-K Top-K 1.02 2.96 0.53 3.41 15.70 2.54 19.50 36.50
R1Sim Top-K 1.05 3.16 0.56 3.36 15.50 2.23 16.30 36.60
CAT-K Auto-labeler 1.19 3.02 0.67 3.21 15.50 2.57 20.30 36.70
CAT-K CRAFT 0.92 2.20 0.69 2.46 10.80 2.26 16.60 26.60
SMART CRAFT 0.91 2.20 0.72 2.60 11.90 2.06 17.10 25.90
CAT-K collision P-Sc. 15.70 → 10.80   (31.2% reduction)
CAT-K traffic P-Sc. 36.50 → 26.60   (27.1% reduction)
SMART traffic P-Sc. 38.80 → 25.90   (33.2% reduction)

Test-time alignment analysis

CPE identifies risky action distributions before failures materialize and mitigates long-horizon error accumulation.

This proactive signal enables test-time correction before execution, whereas rule-based evaluators usually react only after explicit violations have occurred.

Abnormal-agent warning rate over different warning lead times
Proactive risk assessment. CPE warns 48.7% of abnormal behaviors at least 1.0 s in advance and 17.7% at least 2.0 s in advance.
Abnormal-agent rate curves of CRAFT and CAT-K over closed-loop rollout time
Error accumulation mitigation. At the 9 s horizon, CRAFT achieves a 23.79% lower abnormal-agent rate than CAT-K.

Robustness

Training scale, guidance strength, and candidate size jointly affect test-time alignment.

Increasing the preference-training data consistently improves behavioral metrics, indicating that CPE benefits from broader coverage of simulator-induced failure modes. At inference time, the guidance scale β controls the strength of calibration: a small β provides limited correction, while an overly large β over-calibrates the base distribution and degrades performance. Increasing the candidate pool from 16 to 32 improves guidance quality, but a larger pool introduces noisier candidates and reduces stability.

Relative performance under different preference-data sizes, guidance scales, and candidate-pool sizes

Contributions

From incomplete log imitation to complete-context preference alignment.

Local-to-global mismatch

Imitation learning on ego-centric logs leads to incomplete context-action associations that degrade closed-loop behavior.

Contextual Preference Evaluator

CPE learns complete-context behavioral preferences from simulator-induced failures grounded by human-aligned driving priors.

Plug-in test-time guidance

CPE is deployed as a plug-in test-time guidance module for autoregressive decoding, improving behavioral rationality while preserving competitive realism.

Citation

BibTeX

@inproceedings{anonymous2026craft,
  title     = {Bridging Local Observation and Global Simulation in Closed-Loop Traffic Modeling},
  author    = {Ziyan Wang and Tan Xiang and Peng Chen and Xintao Yan},
  booktitle = {},
  year      = {2026}
}