Retrieval-Augmented Large Language Model Decision-Making for Autonomous Driving Guided by Chinese Philosophical Wisdom

Xiaojun Bi; Xiaoyuan Ma; Yiwen Sun; Tianren Huang; Chaoran Liu; Bokai Huang; Hao Yang; Baichuan Mo

cs.AIOct 2, 2026

Retrieval-Augmented Large Language Model Decision-Making for Autonomous Driving Guided by Chinese Philosophical Wisdom

Authors: Xiaojun Bi, Xiaoyuan Ma, Yiwen Sun, Tianren Huang, Chaoran Liu, Bokai Huang, Hao Yang, Baichuan Mo

Organizations: Key Laboratory of Ethnic Language Intelligent Analysis and Security Governance of MOE, and School of Information and Engineering, Minzu University of China, Beijing 100081, China · Institute for Artificial Intelligence, Peking University, Beijing 100871, China · State Key Laboratory of General Artificial Intelligence, BIGAI, Beijing 100080, China · School of Electronics Engineering and Computer Science, Peking University, Beijing 100871, China · School of Earth and Space Sciences, Peking University, Beijing 100871, China · Yuanpei College, Peking University, Beijing 100871, China · Department of Civil Engineering, Tsinghua University, Beijing 100084, China

Abstract

Autonomous driving decision systems must balance safety, efficiency, and social norms in complex traffic interactions. Philosophical and ethical considerations have received limited attention in existing autonomous driving decision-making approaches based on numerical optimization, sequence prediction, and large language models (LLMs). We propose Chinese Philosophical Wisdom-Guided Driving (CPW-Drive), a closed-loop retrieval-augmented generation (RAG) framework that incorporates value guidance derived from Chinese philosophy into autonomous driving decision-making. Using Chinese Confucian thought as its knowledge source, CPW-Drive consolidates LLM-extracted keywords from relevant classical texts into driving-relevant value principles through manual screening and validation. It then contextualizes these principles through scenario-specific cases to form retrievable and reusable value guidance. We further propose Physics-aware Spatial Similarity Retrieval (PSSR), which compares vehicle layouts and velocity-extrapolated states to retrieve physically relevant historical cases. On Highway-env's multilane highway-driving task, CPW-Drive achieves success rates of 93.0%, 86.0%, and 72.0% across three traffic configurations. These results outperform the strongest baseline by 8.0, 22.5, and 25.0 percentage points, respectively. Across all configurations, CPW-Drive achieves the highest collision-free step count and maintains a low lane-change frequency. The results suggest that structured value guidance can improve simulated closed-loop safety and stability while introducing efficiency and latency trade-offs.

Figures & tables

Figure 1: Conceptual illustration of CPW-Drive. Driving-relevant value principles derived from Chinese philosophical wisdom provide retrievable guidance for closed-loop LLM-based driving decisions.

Figure 2: System overview of CPW-Drive. Driving-relevant value principles are extracted and combined with scenario-specific driving cases to construct the Philosophical Driving Knowledge Base, from which PSSR retrieves relevant historical cases to support closed-loop LLM decision-making.

	4 lanes, density = 2.0		5 lanes, density = 2.5		5 lanes, density = 3.0
Model	SR $\uparrow$ (%)	CFS $\uparrow$	SR $\uparrow$ (%)	CFS $\uparrow$	SR $\uparrow$ (%)	CFS $\uparrow$
DriveLikeAHuman	$51.5\pm 5.8$	$19.75\pm 1.25$	$50.5\pm 10.1$	$\underline{24.50\pm 1.07}$	$33.5\pm 5.8$	$12.90\pm 1.74$
DiLu	$75.0\pm 7.5$	$22.90\pm 2.66$	$59.0\pm 5.7$	$22.15\pm 1.34$	$\underline{47.0\pm 5.9}$	$14.24\pm 1.44$
LimSim++	$\underline{85.0\pm 8.8}$	$\underline{27.35\pm 1.98}$	$\underline{63.5\pm 8.8}$	$21.25\pm 2.13$	$37.5\pm 7.9$	$\underline{14.32\pm 2.51}$
CPW-Drive (Ours)	$\mathbf{93.0\pm 5.4}$	$\mathbf{28.12\pm 1.51}$	$\mathbf{86.0\pm 7.0}$	$\mathbf{27.10\pm 1.55}$	$\mathbf{72.0\pm 6.3}$	$\mathbf{22.44\pm 1.37}$

Table 1: Main closed-loop driving performance under different traffic settings. Best and second-best results for SR, CFS, and ES within each setting are shown in bold and underlined, respectively. LCF is reported as a descriptive behavioral metric and is not ranked.

	4 lanes, density = 2.0		5 lanes, density = 2.5		5 lanes, density = 3.0
Method	SR $\uparrow$ (%)	CFS $\uparrow$	SR $\uparrow$ (%)	CFS $\uparrow$	SR $\uparrow$ (%)	CFS $\uparrow$
CPW-Drive	$\mathbf{93.0\pm 5.4}$	$\underline{28.12\pm 1.51}$	$\mathbf{86.0\pm 7.0}$	$\mathbf{27.10\pm 1.55}$	$\mathbf{72.0\pm 6.3}$	$\mathbf{22.44\pm 1.37}$
No RAG	$84.0\pm 4.6$	$25.85\pm 1.22$	$70.0\pm 5.8$	$24.10\pm 1.27$	$57.0\pm 6.7$	$18.05\pm 1.69$
Text RAG	$84.5\pm 6.0$	$25.55\pm 1.82$	$74.5\pm 5.5$	$24.70\pm 1.45$	$61.0\pm 7.7$	$18.60\pm 1.91$
RAG w/o Philosophy	$85.5\pm 5.0$	$26.45\pm 1.08$	$77.0\pm 6.3$	$25.25\pm 1.32$	$67.0\pm 5.9$	$20.35\pm 1.78$
Philosophy in Prompt	$\underline{92.0\pm 4.8}$	$27.82\pm 1.57$	$\underline{82.0\pm 5.4}$	$\underline{26.40\pm 1.27}$	$\underline{69.5\pm 6.4}$	$\underline{21.05\pm 1.67}$

Table 2: Ablation results under different traffic settings. Best and second-best results for SR, CFS, and ES within each setting are shown in bold and underlined, respectively. LCF is reported as a descriptive behavioral metric and is not ranked.

Figure 3: Safe-step survival curves under different database sizes and Top- $k$ settings. Panels (a)–(c) correspond to 50, 90, and 135 historical cases, respectively.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: System and user prompts used by CPW-Drive for closed-loop driving decision-making.

Figure 5: Example of a current driving scenario and one representative historical case retrieved by PSSR.

Figure 6: LLM decision response generated from the current driving scenario and the retrieved historical cases. The response provides the reasoning process and ends with the action ID in the required output format.

Figure 7: System prompt used for the Philosophy in Prompt setting.

Figure 8: System prompt used for the Conservative Prompt setting.

Figure 9: Distribution of PICACO-based Confucian value evaluation scores for CPW-Drive. Each circle represents one of the 40 sampled decision frames, diamonds denote the mean scores, and the values on the right report the mean $\pm$ standard deviation. The shaded row shows the overall Confucian conformity score obtained by combining the four value dimensions with response relevance.

Figure 10: Qualitative comparison of historical scene retrieval results. (a) Query scenario. (b)–(d) The three highest-ranked cases retrieved by text-vector retrieval. (e)–(g) The three highest-ranked cases retrieved by Physics-aware Spatial Similarity Retrieval.

Method	LLM Inference Time (s) $\downarrow$	Total Time (s) $\downarrow$	P95 Total Time (s) $\downarrow$
DriveLikeAHuman	$\mathbf{5.24\pm 2.23}$	$\mathbf{5.24\pm 2.23}$	$\mathbf{9.77}$
DiLu	$13.55\pm 6.06$	$14.81\pm 6.18$	$24.52$
LimSim++	$39.79\pm 24.43$	$42.30\pm 25.28$	$89.99$
CPW-Drive (Ours)	$20.06\pm 10.05$	$20.06\pm 10.05$	$40.71$

Appendix

Table 3: Online decision latency comparison of different methods. LLM inference time and total time are reported as mean

\pm

standard deviation; P95 denotes the 95th percentile of the total time, and

\downarrow

indicates lower is better.

Metric	DiLu	LimSim++	Local Vector Retrieval	CPW-Drive (Ours)
Retrieval Time (ms) $\downarrow$	$1257.28\pm 1145.53$	$1506.31\pm 1867.74$	$20.96\pm 2.64$	$\mathbf{4.23\pm 1.05}$

Appendix

Table 4: Historical scene retrieval time comparison of different methods. Retrieval time is reported as mean

\pm

standard deviation, and

\downarrow

indicates lower is better.

Explore similar work

May 19, 2026

Reason--Imagine--Act: Closed-Loop LLM Decision Making with World Models for Autonomous Driving

Large language models (LLMs) are promising for autonomous driving, but semantics-only decision policies can yield physically unsafe behavior in dynamic traffic. Existing methods either perform online language reasoning without explicit dynamics verification or use world models mainly in offline pipelines, leaving a gap between semantic intent and physical feasibility at decision time. We propose Reason--Imagine--Act (RIA), a closed-loop framework that couples an LLM reasoner with an action-conditioned world model for online safety verification. At each step, the LLM proposes an action template and candidate sub-actions, the world model performs short-horizon rollouts, and a safety scorer selects the safest executable action with feedback to the next reasoning step. Under a unified CARLA point-goal protocol (1000 episodes), RIA achieves 80.05% route completion, 51.10% arrival rate, and 0.20% collision rate. Under the same closed-loop interface, RIA consistently outperforms training-free baselines, including CARLA TM and MADA, on core closed-loop metrics. For reproducibility, code is available at https://github.com/pku-smart-city/source_code/tree/main/RIA.

Zhengqi Sun, Yiwen Sun, Boxuan Liu +3

Department of Information Management, Peking University, Beijing 100871, China · School of Intelligence Science and Technology, Peking University, Beijing 100871, China · State Key Laboratory of General Artificial Intelligence, BIGAI, Beijing 100080, China +3

Language Model-Based Control World Models World Models for Autonomous Driving Autonomous Driving Planning Autonomous Driving Simulation

Ask PDF

Jun 7, 2026

LUNA-AD: Lightweight Uncertainty-Aware Language Model with Lifelong Learning for Autonomous Driving

While large language models (LLMs) offer promising reasoning capabilities, their integration into safety-critical driving systems is hindered by limited reasoning diversity, high computational overhead, and static learning paradigms. To address these challenges, we propose LUNA-AD, a lightweight uncertainty-aware language model with lifelong learning for autonomous driving (AD). LUNA-AD features a tri-system architecture that reconciles complex multimodal behavioral reasoning, efficient deployment, and continual refinement. We design a multi-agent analytical system to generate uncertainty-aware decision-making demonstrations through diverse hypothesis exploration. A dual-head lightweight heuristic model is distilled to unify the inference of decision distributions and textual explanations while enabling efficient deployment. Furthermore, a reflection-driven lifelong learning mechanism operates on multimodal decision outputs and preserves strategic diversity, allowing for the refinement of candidate decisions and rationales via closed-loop feedback to enhance driving robustness. Extensive experiments on nuPlan benchmarks demonstrate that LUNA-AD achieves state-of-the-art success rates under both non-reactive and reactive modes, with drastically reduced inference latency compared to existing knowledge-driven AD frameworks.

Ruoyu Yao, Pei Liu, Ruiguo Zhong +3

The Hong Kong University of Science and Technology (Guangzhou), China

Continual Learning for LLMs Language Model-Based Control Continual Learning Autonomous Driving Planning Decision-Making under Uncertainty

Ask PDF

Mar 31, 2026

C-TRAIL: A Commonsense World Framework for Trajectory Planning in Autonomous Driving

Trajectory planning for autonomous driving increasingly leverages large language models (LLMs) for commonsense reasoning, yet LLM outputs are inherently unreliable, posing risks in safety-critical applications. We propose C-TRAIL, a framework built on a Commonsense World that couples LLM-derived commonsense with a trust mechanism to guide trajectory planning. C-TRAIL operates through a closed-loop Recall, Plan, and Update cycle: the Recall module queries an LLM for semantic relations and quantifies their reliability via a dual-trust mechanism; the Plan module injects trust-weighted commonsense into Monte Carlo Tree Search (MCTS) through a Dirichlet trust policy; and the Update module adaptively refines trust scores and policy parameters from environmental feedback. Experiments on four simulated scenarios in Highway-env and two real-world levelXData datasets (highD, rounD) show that C-TRAIL consistently outperforms state-of-the-art baselines, reducing ADE by 40.2%, FDE by 51.7%, and improving SR by 16.9 percentage points on average. The source code is available at https://github.com/ZhihongCui/CTRAIL.

Zhihong Cui, Haoran Tang, Tianyi Li +4

the Department of Informatics, University of Oslo, Norway · the Department of Computing, Hong Kong Polytechnic University, Hong Kong, China · the Department of Computer Science, Aalborg University, Denmark

Autonomous Driving Planning Autonomous Driving

Ask PDF