ReactSim-Bench: Benchmarking Reactive Behavior World Model Simulation in Autonomous Driving

arXiv 2026 59.1 benchmark, application

TLDR

Introduces ReactSim-Bench to evaluate reactive capability of behavior world models in autonomous driving using decoupled control and diverse metrics.

Reasoning

Strengths include a novel benchmark that decouples agent and AV control, systematic evaluation of multiple model architectures, and practical metrics for safety and rule compliance. Weaknesses are the narrow focus on autonomous driving and potential lack of generalizability to other domains.

Read-first score

Read-first score 59.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 39.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=analysis,benchmark,evaluation,metric

Citation impact 18%

88.9

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.88920254

Topical relevance 29%

55.7

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 349.

Keyword Scores

world model

world simulator

interactive world model

generative world model

world dynamics prediction

video world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Introduces ReactSim-Bench, the first benchmark specifically designed to evaluate reactive capability of behavior world model simulators in autonomous driving.
Decouples control of agents and the AV, using AV behaviors that differ from the log to test agent responses.
Constructs a pipeline that uses an AV planner model to generate candidate behaviors, filtered by rules and manual verification.
Proposes collision metrics, map-based metrics, and kinematic feasibility metrics to evaluate safety and rule compliance of reactive responses.
Provides a systematic evaluation of state-of-the-art models across Transformer-based, diffusion-based, and next-token-prediction-based architectures.

Methodology

The benchmark decouples agent and AV control by feeding the simulator AV behaviors that differ from the log, generated by an AV planner model and filtered via rules and manual verification. It constructs 2,636 test scenarios across three categories and evaluates reactive responses using collision, map-based, and kinematic feasibility metrics. The evaluation compares multiple model architectures and analyzes the effect of replan frequency.

Key Results

Systematic evaluation of state-of-the-art models across Transformer, diffusion, and next-token-prediction architectures reveals performance differences in reactive capability. Analysis of replan frequency provides insights into how often models need to update their predictions to maintain safe and feasible responses.

Limitations

The benchmark is limited to 2,636 scenarios across three categories, which may not fully represent the diversity of real-world reactive driving situations.
Manual verification in the pipeline introduces potential subjectivity and scalability issues.
The AV planner model used to generate candidate behaviors may not cover all possible AV behaviors, limiting the range of reactive challenges tested.

Resources

Paper