RESPOND: Risk-Enhanced Structured Pattern for LLM-driven Online Node-level Decision-making in Autonomous Driving

Current LLM-based driving agents relying on unstructured plain-text memory suffer from low-precision scene retrieval and inefficient reflection. To address this, we present RESPOND, a structured decision-making framework for LLM agents grounded in risk patterns. RESPOND constructs a unified 5 × 3 ego-centric matrix that encodes spatial topology and road constraints, enabling consistent retrieval of spatial–risk configurations. Building on this, a hybrid Rule+LLM pipeline employs a two-tier memory lookup: exact patterns ensure rapid, safe action reuse in high-risk contexts, while sub-patterns facilitate personalized style adaptation under low risk. Furthermore, a pattern-aware reflection mechanism abstracts tactical corrections from crash frames to update structured memory, achieving "one-crash-to-generalize" learning. Experimental results demonstrate the effectiveness of this design. In highway-env, RESPOND surpasses state-of-the-art LLM-based and RL-based agents while generating substantially fewer collisions. With step-wise human feedback, the agent acquires a Sporty driving style within approximately 20 decision steps through sub-pattern abstraction. For real-world validation, we evaluate RESPOND on 53 high-risk cut-in scenarios extracted from the HighD dataset. For each event, we intervene at the moment immediately preceding the cut-in and allow RESPOND to re-decide the driving action. Compared to the recorded human behavior, RESPOND’s decisions reduce subsequent risk in 84.9% of these scenarios, demonstrating the feasibility and practical relevance of RESPOND under real-world driving conditions. These results highlight the potential for real-world autonomous driving, personalized driving assistance, and proactive hazard mitigation. Code will be released at: https://github.com/gisgrid/RESPOND .