acceptodds
Withdrawn from ICLR 2027

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

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

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.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Autonomous driving, Large language models, Chinese philosophical wisdom, Confucian thought, Physics-aware spatial similarity retrieval

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_RG5nOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes CPW-Drive, an LLM-based driving framework guided by Confucian values. It turns these values into manually reviewed driving cases and retrieves relevant cases using a physics-aware similarity measure. Experiments report improved safety in Highway-env, at the cost of driving efficiency.

Critical strengths

The paper takes an imaginative approach to driving decision-making. Connecting Confucian thought with scenario-based guidance offers a distinctive perspective.

Critical weaknesses
  1. The evaluation relies entirely on Highway-env, which is too simplified to convincingly validate the proposed autonomous-driving method. Testing is further restricted to highway-v0, without covering merging or roundabouts. My concern is both the simulator’s limited realism and the narrow task coverage. Evaluation in a more realistic driving simulator is needed to support the paper’s claims about complex traffic interactions.

  2. The method relies on retrieving relevant examples from a small, curated case library. Its performance on substantially different scenarios is therefore an important concern. Using separate random seeds provides some evidence on new instances of the same task. However, the paper does not establish generalization to unseen road layouts, interaction types, or clearly held-out traffic configurations.

  3. The anonymous repository does not provide the case data and runnable configurations for key controls such as RAG w/o Philosophy. It is also unclear whether these cases received the same manual review and correction as the philosophical cases. These omissions make the reported comparisons difficult to verify and weaken confidence that the gains can be attributed to philosophical guidance.

Questions for authors
  1. How were the 135 cases selected from more than 3,000 candidate scenarios? What criteria were used to revise or reject their reasoning and driving actions?

  2. Were the non-philosophical cases independently generated and then manually reviewed under the same standards as the philosophical cases? Please clarify the differences in case construction and inference prompts, and release the corresponding data and code.

  3. Which lane configurations and traffic densities were used to construct the case library? Were any test configurations entirely held out? Beyond using different random seeds, how was overlap between library and test scenarios assessed?

  4. Which exact prompts, configurations, and seed sets produced the reported tables? The released entry point runs one round of 20 seeds per configuration, while the paper reports ten rounds. Please provide the complete evaluation procedure.

  5. See weaknesses.

Reviewer_943LOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CPW-Drive, a retrieval-augmented LLM framework that incorporates principles from Confucian philosophy into autonomous driving decisions. The authors derive ten driving-relevant principles from 131 classical texts and use them to construct a knowledge base of 135 curated driving cases, each containing a scenario, relevant values, reasoning, and an action. They also introduce Physics-aware Spatial Similarity Retrieval (PSSR), which retrieves relevant cases by considering vehicle positions and short-term motion rather than relying solely on textual similarity.

The approach is evaluated in the Highway-env simulator across three traffic configurations and compared against several existing LLM-based driving methods. CPW-Drive achieves higher success rates and more collision-free steps across all three settings, although with lower driving efficiency. The authors also conduct ablation studies to examine the contributions of philosophical guidance and the retrieval mechanism, along with additional analyses of retrieval efficiency and inference latency.

Critical strengths

(1) One of the main strengths of the paper is the proposed PSSR retrieval method. Rather than relying on text similarity, it takes into account vehicle positions and short-term motion to identify relevant driving scenarios. The ablation results show consistent improvements over text-based retrieval across all three traffic settings, which provides convincing evidence that this component contributes to the overall performance. I find this particularly valuable because the approach is relatively simple, interpretable, and well suited to the driving task.

(2) Another strength is the way the authors incorporate philosophical principles into concrete driving examples instead of simply including them in the prompt. This makes the proposed framework more interesting than a straightforward prompting approach. The experiments also suggest that using scenario-specific guidance can achieve a better balance between safety and driving efficiency than directly injecting the philosophical principles into the prompt. I think this is a worthwhile contribution, particularly for exploring how higher-level values can be incorporated into LLM-based decision-making.

Critical weaknesses

(1) My main concern is that the paper does not clearly establish how much of the performance improvement comes specifically from the Confucian philosophical principles. In particular, the Philosophy in Prompt baseline achieves results quite close to those of the full framework, which makes it difficult to assess the added benefit of the proposed approach. While the ablation studies are useful, I think a stronger comparison with non-philosophical, safety-oriented guidance would help clarify whether the improvements are really due to the philosophical component or mainly to the retrieval mechanism and curated driving examples. This is important because the integration of Confucian philosophy is one of the paper's central contributions.

(2) Another weakness is the limited experimental setting. The evaluation is conducted in a single simulator, with relatively short driving episodes and a small number of traffic configurations. It is therefore difficult to judge how well the approach would generalize to more complex and realistic driving situations. Additionally, the reported inference time of around 20 seconds per decision is a significant limitation for real-time autonomous driving. Although the authors acknowledge this issue, it still raises concerns about the practical applicability of the proposed framework.

Questions for authors

(1) The Philosophy in Prompt baseline achieves results fairly close to those of the full CPW-Drive framework. Could the authors clarify whether these differences are statistically significant? I would also be interested in understanding whether similar improvements could be achieved using more general safety-oriented principles rather than Confucian philosophy. This would help clarify the contribution of the philosophical component, which is central to the paper.

(2) The reported average inference time is around 20 seconds per decision, while the simulator operates at a policy frequency of 1 Hz. Could the authors explain how this difference is handled during evaluation? In particular, does the simulator pause while waiting for the LLM response? I think this is important for interpreting the reported safety improvements.

(3) Regarding PSSR, how much of the improvement comes from incorporating the velocity-extrapolated states rather than simply comparing current vehicle positions? A comparison with a spatial-only retrieval variant would help assess the importance of this particular design choice.

(4) Appendix E mentions that the retrieval databases used by DiLu and LimSim++ contain 20 historical examples, compared with 135 for CPW-Drive. Could the authors clarify how much this difference in database size and curation might influence the results? I would find a comparison under similar conditions useful for judging the improvements attributed to the proposed method.

Reviewer_amJ8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CPW-Drive, a retrieval-augmented LLM framework that incorporates Confucian values into closed-loop decision-making for autonomous highway driving. The authors first extract driving-relevant value principles from Confucian classics and construct a historical driving case library using traffic scenarios sampled from Highway-env. During inference, the framework retrieves similar historical cases using a physics-aware similarity measure based on Gaussian kernels and constant-velocity extrapolation. The retrieved cases, together with Confucian value principles, are then used to guide the LLM in making driving decisions.

Critical strengths
  1. The proposed approach is interesting, offering a novel perspective on incorporating Confucian philosophical values into autonomous driving decision-making.

  2. The method achieves improved performance over several baselines in Highway-env. The ablation studies also provide evidence for the effectiveness of RAG and Confucian value guidance.

Critical weaknesses
  1. Limited evaluation scenarios. The method is evaluated only in the highly simplified Highway-env simulator, which cannot adequately reflect the complexity of real-world driving. Evaluation on more realistic driving benchmarks or real-world driving datasets is needed to demonstrate its practical value and generalizability.

  2. Insufficient evidence for the effectiveness of philosophical guidance. The conclusions regarding RAG and Confucian values are primarily supported by experiments in a few Highway-env configurations. It remains unclear whether the improvements specifically come from Confucian philosophical principles or general safe-driving guidance. More rigorous comparisons with standard traffic safety principles are needed.

  3. Poor real-time performance. The framework requires approximately 20 seconds per decision, which is significantly slower than the 1 Hz decision frequency used in the simulation. This substantial inference latency raises concerns about the feasibility of deploying the method in practical autonomous driving systems.

  4. Limited technical novelty and insufficient retrieval baselines. The proposed PSSR primarily relies on Gaussian-kernel spatial similarity and constant-velocity extrapolation. Its technical novelty is relatively limited, and the experiments lack comparisons with simpler physics-based retrieval methods, such as relative-distance or time-to-collision-based similarity measures.

  5. Limited validation of the constructed knowledge base. The historical case library contains only 135 manually screened cases, which may not adequately cover diverse driving situations. Moreover, the paper provides limited quantitative evidence regarding the quality, consistency, and reliability of the generated reasoning and driving decisions.

Questions for authors

N/A

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.