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

Chen, Dan; Huang, Heye; Chen, Tiantian; Li, Zheng; Li, Yongji; Xu, Yuhui; Chen, Sikai · 2025 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2512.20179

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the limitations of current Large Language Model (LLM)-based driving agents, which rely on unstructured plain-text memory. Such reliance leads to low-precision scene retrieval, inefficient reflection learning, and an inability to encode spatial-risk dependencies critical for safety. To solve this, the authors propose RESPOND, a structured decision-making framework grounded in risk patterns. The system aims to enable precise retrieval of spatial-risk configurations, efficient reuse of safe actions, and rapid generalization from sparse safety-critical events. The methodology centers on a unified 5 × 3 ego-centric risk matrix derived from the Driver Risk Field (DRF) model. This matrix discretizes spatial topology and road constraints into a compact representation, replacing raw sensor data or natural language descriptions. RESPOND employs a two-tier structured memory system: Layer 1 stores exact global risk patterns for high-precision retrieval, while Layer 2 stores abstracted sub-patterns (e.g., front, rear, left, right risks) to handle data sparsity and support tactical reasoning. A hybrid Rule+LLM decision pipeline utilizes this memory; in high-risk contexts, it prioritizes exact pattern reuse or sub-pattern-constrained rule-based actions to ensure safety, bypassing LLM reasoning when possible. In low-risk contexts, it allows for personalized style adaptation. Additionally, a pattern-aware reflection mechanism analyzes pre- and post-crash frames to extract tactical corrections, updating the structured memory to achieve "one-crash-to-generalize" learning. Experimental results demonstrate RESPOND’s effectiveness in both simulation and real-world scenarios. In the highway-env simulator, RESPOND outperformed state-of-the-art LLM-based and reinforcement learning agents, generating substantially fewer collisions. The system also demonstrated efficient personalization, acquiring a "Sporty" driving style within approximately 20 decision steps using step-wise human feedback and sub-pattern abstraction. For real-world validation, the authors evaluated RESPOND on 53 high-risk cut-in scenarios from the HighD dataset. By intervening immediately before the cut-in and allowing RESPOND to re-decide the action, the system reduced subsequent risk in 84.9% of cases compared to the recorded human behavior. The significance of this work lies in its demonstration that structured, risk-aware memory can overcome the semantic drift and retrieval inefficiencies of text-based LLM agents. By bridging quantitative risk modeling with symbolic reasoning, RESPOND offers a feasible approach for real-world autonomous driving, personalized driving assistance, and proactive hazard mitigation. The framework provides a scalable method for accumulating safety knowledge from limited failure data, enhancing the reliability and interpretability of LLM-driven decision systems in complex traffic environments.

Key finding

The RESPOND framework reduces subsequent risk in 84.9% of high-risk cut-in scenarios compared to human drivers and outperforms existing LLM and reinforcement learning agents in simulation with fewer collisions.

Methodology

simulation_modeling

Sample size: 53

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via author_sweep_intake on 2026-05-28.

StageOutcomeToolModelPromptAttemptsCompleted
discover success author_sweep 2 2026-05-28
archive success canonical_url 1 2026-06-04
extract success cached 3 2026-06-10
clean success clean 1 2026-06-04
chunk success chunk 1 2026-06-04
embed success embed Qwen/Qwen3-Embedding-8B 1 2026-06-04
enrich success 1 2026-05-28
promote success 1 2026-06-04
summarize success llm qwen3.6-27b-prismaquant summ-v5 2 2026-06-10
tag success vector_similarity 15 2026-06-11
verify success 2 2026-06-10

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).