Preprint

EvoDrive: Pareto Evolution for Safety-Critical Autonomous Driving via Self-Improving LLM Agents

Simulator-grounded LLM agents evolve adversarial driving scenario generators while preserving realism through a Pareto archive.

arXiv Code (Coming soon) Video

Tong Nie^1,2,†, Yuewen Mei^2,†, Yihong Tang^3,4, Junlin He¹, Jie Deng¹, Jian Sun^2,✉, Wei Ma^1,✉

¹The Hong Kong Polytechnic University ²Tongji University ³McGill University ⁴Mila-Quebec AI Institute

^† Equal contribution. ^✉ Corresponding authors.

Abstract

Agentic evolution with simulator evidence.

Generating safety-critical scenarios requires exposing failures while keeping scenes physically plausible. EvoDrive uses a simulator-grounded actor-critic architecture where memory-driven agents propose bounded generator edits, critics filter implausible candidates, and a self-evolving world evaluator routes promising proposals under finite simulation budgets. A Pareto archive preserves diverse attack-realism trade-offs and guides future evolution through closed-loop feedback.

Motivation

Safety validation needs rare, realistic stress cases.

Handcrafted generators stay near known priors.

Handcrafted rules and scalar rewards tend to explore near designer-specified behaviors.

Unconstrained agents exploit simulators.

General LLM agents can find high-attack cases, but without simulator grounding they may exploit unrealistic failures.

The trade-off is multi-objective.

Useful scenarios should expand the attack-realism frontier instead of collapsing to one fixed compromise.

Method

A closed-loop evolution system for scenario generators.

EvoDrive proposes, validates, evaluates, and archives generator edits. Every accepted candidate is backed by simulator labels, while memory and lineage records guide the next proposal.

EvoDrive evolution loop from generator inputs to Pareto archive outputs. — Actor and repair agents propose bounded edits, validators and critics reject invalid candidates, closed-loop simulation supplies labels, and the Pareto archive feeds future evolution.

Proposal

Memory-conditioned actors edit declared generator surfaces rather than arbitrary code.

Verification

Specialized critics and deterministic validators filter malformed or implausible candidates before rollout.

Overview, results, and scenario visualizations.

Results

Consistent frontier expansion across simulators.

EvoDrive improves Pareto area for every MetaDrive generator family and produces large attack-realism gains on CARLA SafeBench.

8 / 8 MetaDrive generators improve PF-Area@3

+255.7% CARLA AT PF-Area@3 gain

0.163 -> 0.106 Average SafeBench collision rate after training

0.854 -> 0.884 Average SafeBench overall score after training

Pareto frontier dynamics across CARLA evolution rounds. — Pareto-frontier dynamics broaden over evolution rounds.

SAC training curves with original and EvoDrive evolved scenarios. — Evolved scenarios improve downstream policy robustness in MetaDrive.

Mechanisms

Evolution discovers structured generator edits.

Structured edit surfaces keep proposals bounded and interpretable.

Example generator evolution trajectory. — Lineage traces expose accepted and rejected proposal paths.

World-guided candidate ranking before and after evolution. — World-guided ranking allocates simulation budget to higher-yield candidates.

Scenario Examples

Evolved scenes are more adversarial while remaining plausible.

Qualitative examples show original and evolved ChatScene scenarios in CARLA, including RGB and bird's-eye-view renderings.

Original and EvoDrive evolved CARLA scenarios from ChatScene. — Evolution places adversarial actors into more informative interaction windows while preserving map consistency.

Citation

Cite EvoDrive

@misc{nie2026evodrive,
  title         = {EvoDrive: Pareto Evolution for Safety-Critical Autonomous Driving via Self-Improving LLM Agents},
  author        = {Tong Nie and Yuewen Mei and Yihong Tang and Junlin He and Jie Deng and Jian Sun and Wei Ma},
  year          = {2026},
  eprint        = {2606.03678},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  doi           = {10.48550/arXiv.2606.03678},
  url           = {https://arxiv.org/abs/2606.03678}
}