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

1 INTRODUCTION

• Autonomous driving decisions must balance safety, efficiency, and traffic etiquette, while numerical optimization and prediction objectives do not explicitly encode every relevant social norm.Data-driven approaches also face interpretability and generalization challenges in open-ended, long-tail scenarios.Source: §1 · pp. 1–2
• The introduction identifies a gap in representing culturally grounded philosophical principles as reusable, scenario-level driving examples.It states that this gap limits systematic application of abstract values to concrete traffic interactions.Source: §1 · p. 2
• CPW-Drive derives ten driving-relevant value principles from Confucian classics through keyword extraction, manual screening, and conceptual integration, then combines them with driving scenarios as retrievable cases.PSSR retrieves historical cases relevant to the current decision.Source: §1 · p. 2
• Across three traffic configurations, closed-loop experiments report higher success rates, more collision-free steps, and restrained driving under Confucian value guidance, with an efficiency trade-off.The experiments compare CPW-Drive across three traffic configurations.Source: §1 · p. 2

2.1 LLMS FOR AUTONOMOUS DRIVING DECISION-MAKING

• LLM driving work spans language-based planning, reasoning, memory, and closed-loop control, while grounding philosophical principles in concrete driving scenarios remains underexplored.Other approaches also retrieve traffic regulations or learn value preferences.Source: §2.1 · pp. 2–3

2.2 ETHICAL AND PHILOSOPHICAL VALUES IN AUTONOMOUS DRIVING

• Ethical driving research includes collision-dilemma studies and rule- or constraint-based methods, but continuous driving use of scenario-grounded philosophical values remains underexplored.CPW-Drive represents Confucian values as natural-language principles and historical cases rather than scalar rewards or optimization objectives.Source: §2.2 · p. 3

2.3 RETRIEVAL-AUGMENTED DECISION-MAKING AND SCENE SIMILARITY MATCHING

• Existing retrieval methods use language descriptions, embeddings, or structured scene representations, which may not directly measure similarity over vehicle physical states.PSSR jointly considers current spatial configuration and a short-term future spatial state estimated from vehicle velocities.Source: §2.3 · pp. 3–4

3.1 LLM-BASED CLOSED-LOOP DRIVING FRAMEWORK

• CPW-Drive loops from scene encoding to LLM reasoning and decision execution, using retrieved cases and available actions before feeding the updated state into the next step.Each case provides a scenario, value principles, reasoning process, and driving action.Source: §3.1 · p. 4

3.2 DERIVATION OF DRIVING-RELEVANT VALUE PRINCIPLES

• LLM-extracted keywords from 131 Confucian classics are screened and integrated into ten driving-relevant value principles, which guide but do not directly prescribe actions.The principles are combined with scenarios, reasoning, and actions to form historical driving cases.Source: §3.2 · pp. 4–5

3.3 PHILOSOPHICAL DRIVING KNOWLEDGE BASE CONSTRUCTION

• The knowledge base stores cases as C = (S, V, R, A), preserving each traffic scenario, relevant values, reasoning, and resulting action.More than 3,000 candidate Highway-env scenarios were collected, reviewed, and filtered into a final knowledge base of 135 cases.Source: §3.3 · p. 5

3.4 PHYSICS-AWARE SPATIAL SIMILARITY RETRIEVAL

• PSSR compares ego-centered surrounding-vehicle layouts and velocity-extrapolated positions, using anisotropic spatial scales to capture configuration and motion trends.The extrapolation uses a 1.0-second interval; longitudinal and lateral scales reflect their different physical meanings and scales.Source: §3.4 · pp. 5–6
• PSSR sums Gaussian spatial similarities across vehicle pairs, avoiding one-to-one vehicle matching, and normalizes overlap to compare scenes with different vehicle counts.The within-scene self-overlap includes cross terms between distinct vehicles, so it is not simply the number of vehicles.Source: §3.4 · p. 6
• The final similarity weights current-space overlap at 0.7 and extrapolated-space overlap at 0.3, a heuristic setting chosen without extensive tuning because samples were limited.The current layout receives greater weight, while extrapolation complements short-term motion trends.Source: §3.4 · pp. 6–7
• PSSR ranks historical scenes by final similarity and retrieves the top three cases as references for the current LLM decision in the main experiments.Source: §3.4 · p. 7

4.1 EXPERIMENTAL SETUP

• Each configuration uses 10 rounds of 20 runs, with 30 decision steps per run; success requires completing all 30 steps without collision.The three settings are four lanes at density 2.0, five lanes at density 2.5, and five lanes at density 3.0; main experiments retrieve three cases from a 135-case database.Source: §4.1 · p. 7

4.2 MAIN PERFORMANCE ANALYSIS

• 8.0, 22.5, and 25.0 percentage points: CPW-Drive’s success rate exceeds the strongest baseline in the three traffic configurations, and it has the highest success rate and collision-free steps in each.Its lane-change frequency remains generally low, while its efficiency score is below the best baseline in all three configurations.Source: §4.2 · pp. 7–8, Table 1 · p. 8

4.3 ABLATION STUDY

• Removing retrieval lowers success rate and collision-free steps and increases lane-change frequency; replacing PSSR with text-vector retrieval also lowers success rate and collision-free steps.These comparisons support roles for historical cases and physics-aware retrieval of cases matching spatial relations and motion states.Source: §4.3 · p. 8
• Without philosophical content, success rate and collision-free steps fall while efficiency score and lane-change frequency rise; direct value prompting approaches full-model safety but further lowers efficiency and lane changes.Scenario-grounding mitigates the conservatism and efficiency loss of direct value injection, while generic conservative prompting alone does not explain CPW-Drive’s performance.Source: §4.3 · p. 8
• Survival performance improves as the historical-case database grows, and generally improves with higher Top-k; Top-3 and Top-5 have similar final performance, while Top-5 is higher earlier.Figure 3 compares databases of 50, 90, and 135 cases under Top-1, Top-3, and Top-5 retrieval.Source: §4.3 · pp. 8–9, Figure 3 · p. 9

5 CONCLUSION

• Across three simulated traffic configurations, CPW-Drive achieves the highest success rate and collision-free steps, maintains relatively low lane-change frequency, and exhibits lower efficiency score.PSSR retrieves historical cases using vehicle spatial structure and motion trends.Source: §5 · p. 9
• The study is limited to a knowledge base of 135 historical cases and relatively short 30-step trials.The authors identify evaluation in more diverse driving environments and with larger knowledge bases as future work.Source: §5 · pp. 9–10
• Mean end-to-end decision latency is approximately 20 seconds and does not yet meet real-time control requirements.Source: §5 · pp. 9–10

AI USE STATEMENT

• The authors manually screened, revised, and validated the extracted value principles and generated driving cases, and reviewed AI-assisted suggestions and edits.DeepSeek-V4-Pro supported keyword extraction, case generation, and closed-loop driving decisions.Source: AI USE STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

• The public code repository contains implementations, prompts, the historical-case database, configurations, protocols, parameters, and evaluation scripts.Source: REPRODUCIBILITY STATEMENT · p. 10

A CONFUCIAN CLASSICAL WORKS

• 131 Confucian and related classical works are listed as sources for deriving driving-relevant value principles.Source: A CONFUCIAN CLASSICAL WORKS · p. 13
• Each entry gives the title in Hanyu Pinyin followed by its original Chinese title.Source: A CONFUCIAN CLASSICAL WORKS · p. 13
• English titles appear where well established; other English renderings are concise descriptive glosses, not claims of complete translations.Source: A CONFUCIAN CLASSICAL WORKS · p. 13

B PROMPT TEMPLATES AND CLOSED-LOOP DECISION EXAMPLE

• The system prompt defines the driving assistant’s role, value-guided reasoning requirements, and output format, while the user prompt supplies the scene, three PSSR cases, and available actions.Figure 4 presents these as system and user prompts.Source: B PROMPT TEMPLATES AND CLOSED-LOOP… · p. 15, Figure 4 · p. 16
• Figures 5 and 6 trace one decision step from a current scenario and representative PSSR-retrieved case to an LLM response with reasoning and a final action ID.Figure 5 shows the scenario and case; Figure 6 shows the corresponding response.Source: B PROMPT TEMPLATES AND CLOSED-LOOP… · p. 15, Figure 5 · p. 16, Figure 6 · p. 17
• Figures 7 and 8 show the system prompts used for the Philosophy in Prompt and Conservative Prompt ablation settings, respectively.Source: B PROMPT TEMPLATES AND CLOSED-LOOP… · p. 17, Figure 7 · p. 17, Figure 8 · p. 17

C EXTERNAL EVALUATION OF CONFUCIAN VALUE CONFORMITY

• The evaluation used PICACO’s publicly available Confucian-value module and official code without modifying its prompts, definitions, scoring criteria, or aggregation procedure.The authors did not use PICACO’s meta-instruction optimization method.Source: C EXTERNAL EVALUATION OF CONFUCIAN… · pp. 17–18
• One frame was randomly sampled from each of 40 completed runs to reduce temporal redundancy, and PICACO assessed the original query together with the model’s reasoning and decision.Source: C EXTERNAL EVALUATION OF CONFUCIAN… · p. 18
• PICACO evaluated responses on Benevolence, Conformity, Tradition, Security, and query relevance, using 1–5 scores and a combined value-and-relevance conformity score.The overall score averages the four value dimensions, multiplies by relevance, and divides by the maximum score of 5.Source: C EXTERNAL EVALUATION OF CONFUCIAN… · p. 18
• 4.40 ± 0.60 was CPW-Drive’s overall Confucian conformity score across 40 sampled responses.Scores were 3.95 ± 0.68 for Benevolence, 4.68 ± 0.73 for Conformity, 4.60 ± 0.59 for Tradition, 5.00 ± 0.00 for Security, and 4.83 ± 0.45 for relevance.Source: C EXTERNAL EVALUATION OF CONFUCIAN… · pp. 18–19, Figure 9 · p. 18

D QUALITATIVE RETRIEVAL CASE ANALYSIS

• Text-vector retrieval returned cases with the query’s vehicle count but did not consistently preserve longitudinal positions or lane relations, as Figure 10(b)–(d) illustrates.The fixed text templates can yield similar descriptions despite different physical configurations, and the retrieval measure does not explicitly model ego-centered geometry, lane relations, distances, or motion.Source: D QUALITATIVE RETRIEVAL CASE ANALYSIS · p. 19, Figure 10 · p. 19
• PSSR’s Figure 10(e)–(g) cases better preserve principal vehicle relations and lane constraints by comparing overall ego-centered spatial configurations without requiring identical vehicle sets or one-to-one correspondences.The query scenario includes same-lane front, left-lane front, and right-lane rear vehicles.Source: D QUALITATIVE RETRIEVAL CASE ANALYSIS · p. 19, Figure 10 · p. 19
• By combining current and velocity-extrapolated spaces, PSSR accounts for vehicle layout and short-term motion trends when retrieving historical decision references.The comparison distinguishes text-vector retrieval based on linguistic semantics from PSSR’s comparison of vehicle configurations.Source: D QUALITATIVE RETRIEVAL CASE ANALYSIS · pp. 19–20

E COMPUTATIONAL COMPLEXITY AND RUNTIME EFFICIENCY OF CPW-DRIVE

• For M historical cases with at most N surrounding vehicles per scene, PSSR’s online retrieval time is O(MN^2 + M log M) and additional working memory is O(M + N).The time bound includes pairwise spatial similarity and sorting all M scores before selecting the Top-k cases.Source: E COMPUTATIONAL COMPLEXITY AND RUNTIME… · p. 20
• The latency comparison holds the LLM backend constant, but the methods retain their original prompts and retrieval configurations, so the measurements reflect complete pipelines rather than isolated model throughput.Source: E COMPUTATIONAL COMPLEXITY AND RUNTIME… · p. 20
• CPW-Drive requires 20.06 ± 10.05 s per decision on average, exceeding the 1 Hz simulated policy interval and leading the authors to characterize the implementation as a simulation prototype rather than a real-time controller.The latency measurements compare complete method configurations, and the reported standard deviations indicate substantial wall-clock variability.Source: E COMPUTATIONAL COMPLEXITY AND RUNTIME… · p. 20
• 4.23±1.05 ms is PSSR’s mean retrieval latency, the lowest reported; retrieval contributes only a small fraction of CPW-Drive’s 20.06 s end-to-end mean.The local vector comparison searches the same 135 cases, while the DiLu and LimSim++ retrieval times include remote API network and service variability rather than controlled algorithm comparisons.Source: E COMPUTATIONAL COMPLEXITY AND RUNTIME… · pp. 20–21, Table 4 · p. 21

• The two-dimensional Gaussian-overlap integral separates into longitudinal and lateral integrals whose product gives the complete overlap.Source: F DERIVATION OF THE SPATIAL KERNEL AND… · pp. 21–22
• The main-text spatial kernel omits the common multiplicative constant πσxσy from the Gaussian-overlap expression.Source: F DERIVATION OF THE SPATIAL KERNEL AND… · p. 22
• The common factor πσxσy cancels during normalization, so omitting it does not change normalized similarity or historical-scene rankings.Source: F DERIVATION OF THE SPATIAL KERNEL AND… · p. 22
• The longitudinal scale is empirically set to σx = 5 m, while the lateral scale is determined by the desired kernel decay over one lane width.Source: F DERIVATION OF THE SPATIAL KERNEL AND… · pp. 22–23
• For non-empty scenes, normalized similarity is strictly positive and at most 1, and can be interpreted as cosine similarity in feature space.Source: F DERIVATION OF THE SPATIAL KERNEL AND… · pp. 23–24
• The normalization denominator includes cross terms between distinct vehicles within a scene, so it cannot generally be reduced to the vehicle count.Source: F DERIVATION OF THE SPATIAL KERNEL AND… · p. 24