SafeDrive: Knowledge- and data-driven risk-sensitive decision-making for autonomous vehicles with Large Language Models
DOI: 10.1016/j.aap.2025.108299
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 challenge of ensuring safety and adaptability in autonomous vehicles (AVs) operating in dynamic, high-risk environments, particularly regarding rare "long-tail" events. While data-driven paradigms have improved AV performance in normal scenarios, they suffer from data bias and a lack of interpretability, often failing in unpredictable situations. The authors propose SafeDrive, a knowledge- and data-driven risk-sensitive decision-making framework that integrates Large Language Models (LLMs) with comprehensive risk quantification to enhance AV safety and human-like decision-making. The SafeDrive framework consists of four interconnected modules: a Risk Module, a Memory Module, an LLM-powered Reasoning Module, and a Reflection Module. To address the quantification of coupled risks, the authors develop a dynamic Driver Risk Field (DRF) model. Unlike traditional methods that focus only on the forward-facing half-circle, this model provides omnidirectional (360-degree) risk assessment by accounting for vehicle speed, steering dynamics, and interactions with surrounding road users. The Quantified Perceived Risk (QPR) is calculated by combining subjective probability (via the DRF) with objective event costs. For decision-making, the system uses GPT-4 as the driving agent. It processes real-time risk assessments, retrieves relevant past experiences from a memory database, and employs chain-of-thought reasoning to generate actions. A reflection module iteratively corrects decisions and updates the memory, creating a continuous learning loop. The framework was evaluated on real-world traffic datasets characterized by complex interactions, including highways (HighD), intersections (InD), and roundabouts (RounD). The results demonstrate that SafeDrive achieves a 100% safety rate across these diverse scenarios. Furthermore, the system exhibits strong alignment with human driving behavior, with decision alignment exceeding 85%. The omnidirectional risk model effectively identifies specific risk attributes of surrounding vehicles, allowing for targeted risk warnings and adaptive responses such as deceleration or lane changes in high-conflict situations. The significance of this work lies in establishing a novel paradigm that bridges knowledge-driven insights with adaptive data-driven learning. By integrating interpretable risk quantification with the reasoning capabilities of LLMs, SafeDrive addresses the limitations of black-box algorithms and static risk models. The framework’s ability to maintain high safety standards and replicate human-like behavior in unpredictable, high-risk environments suggests substantial potential for advancing higher-level autonomous driving systems, particularly in managing safety-critical long-tail events.
Key finding
The SafeDrive framework achieved a 100% safety rate and over 85% alignment with human decision-making across diverse real-world traffic scenarios.
Methodology
simulation_modeling
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.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | author_sweep | — | — | 2 | 2026-05-28 |
| archive | success | canonical_url | — | — | 5 | 2026-06-06 |
| 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).
- Theoretical Contribution: computational model