CLOSURE:Closed-LOopSelf-evolving large LangUage model framework forREal-time adversarial scenario generation

The validation of autonomous driving systems (ADS) requires rigorous testing against rare, safety-critical scenarios, yet real-world data collection is hindered by the “curse of rarity”. Traditional optimization-based and data-driven methods often lack realism, controllability, and adaptability. To address this, we introduce CLOSURE, a Closed-Loop Self-Evolving Large Language Model Framework for Real-Time Adversarial Scenario Generation. Unlike existing LLM-driven methods that operate offline or rely on static knowledge, CLOSURE integrates three core innovations: (i) a self-evolving Risk Scenario Library (RSL) initialized with real-world accident precursors (NMVCCS) and continuously expanded with newly discovered failure cases, enabling lifelong risk memory; (ii) a physics-constrained, retrieval-augmented generation mechanism that grounds adversarial intents in real-world crash data while ensuring kinematic feasibility via deterministic validation; (iii) a unified semantic representation that treats both real-world seeds and online-generated cases identically, enabling consistent RAG-based risk assessment. Experiments in SUMO show that CLOSURE increases the collision rate from 0.3 to 30.8 for the rule based ego vehicle and from 0.2 to 31.4 for the learning based ego vehicle. The mean minimum Time-to-Collision (mmTTC) decreases from 3.88 s to 1.20 s for the Ego-I and from 2.85 s to 1.17 s for the Ego-II. Diversity measurements indicate an improvement from 0.316 to 1.019 compared with the rule-based perturbation. Ablation studies confirm that both the Risk Analysis Model and the Risk Scenario Library are essential for strong risk amplification and long tail discovery. CLOSURE provides an effective and scalable framework for generating rare, realistic, and diverse safety critical scenarios for the robustness evaluation of autonomous driving systems.