acceptodds
Under review as a conference paper at ICLR 2027

Neuro-Symbolic Drive: Rule-Grounded Faithful Reasoning for Driving VLAs

Abstract

Driving VLA models incorporating Chain-of-Thought (CoT) reasoning are attractive because they leverage pretrained VLM representations and expose intermediate decisions in natural language, yet current rationales often lack the step-by-step decision semantics needed to keep the rationale causally connected to the planned motion. We introduce Neuro-Symbolic Drive, a neuro-symbolic driving framework that supervises a driving VLA with rule-grounded reasoning traces extracted directly from classical rule-based planners. Our key observation is that rule-based planners are symbolic AI systems that already function as executable reasoning engines: they reason about active safety constraints, search over candidate maneuvers, and select a final trajectory. We instrument these planners in simulation to capture both the executed trajectory and the internal decision trace at each rule-evaluation step. Each trace is serialized into structured rule-grounded reasoning and paired with the trajectory to fine-tune Qwen3.5-4B as a driving VLA. Because these traces are derived directly from the planner states that determine the action, they ensure reasoning is structurally coupled to motion generation by construction, rather than by post-hoc alignment. On our simulator-generated benchmark, detailed rule-grounded reasoning reduces ADE@3s from 0.47 to 0.26 and miss rate from 8.30% to 6.40% under three-camera perception, and from 0.54 to 0.26 and 10.13% to 5.99% under eight-camera perception. Neuro-Symbolic Drive thus converts neuro-symbolic planning logic into structured supervision. All source code will be made publicly available upon acceptance.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Vision-Language-Action Models, Autonomous Driving, Neuro-Symbolic AI, Chain-of-Thought Reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_bFVVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes NEURO-SYMBOLIC DRIVE, which extracts reasoning from the execution traces of rule-based planners, including IDM, IDM MOBIL Planner, and PDM-Closed. It maps these traces to a unified four-slot schema covering scenes, constraints, candidate actions, and decisions, and uses them to jointly supervise reasoning and trajectory prediction in vision-language driving models. The authors evaluate different reasoning supervision formats and teacher selection strategies using Qwen3.5-4B on a dataset built from 520 nuPlan scenarios. The results show that, compared with trajectory-only supervision, planner-generated reasoning improves open-loop trajectory prediction under both three-camera and eight-camera settings.

Critical strengths
  1. Motivation and supervision design. The paper addresses the potential mismatch between the language explanations generated by driving models and their actual planning behavior. Instead of generating explanations post hoc, it extracts reasoning from the execution of the rule-based planner that produces the target trajectory. This design is well motivated and provides a reasonable way to link reasoning supervision with planning behavior.

  2. Consistent empirical improvements. Reasoning supervision improves trajectory prediction under both three-camera and eight-camera settings. Detailed reasoning reduces ADE@3s from 0.47 to 0.26 and from 0.54 to 0.26, respectively, while reducing MR@3s from 8.30% to 6.40% and from 10.13% to 5.99%. The multi-planner ablation also shows that combining different teacher planners achieves better overall results than using a single teacher.

Critical weaknesses
  1. Insufficient validation of the core reasoning contribution and faithfulness. The paper's central claim is that planner execution traces provide faithful reasoning. However, the fact that the teacher's execution state and target trajectory come from the same planner execution does not guarantee that the model's generated reasoning remains faithful to its predicted behavior after supervised fine-tuning. The experiments mainly evaluate trajectory prediction errors and do not directly measure the consistency between generated reasoning and teacher traces in terms of scene understanding, constraint assessment, candidate filtering, and final decisions. Figure 3 also shows differences between the GT reasoning and predicted reasoning in scene classification and state descriptions. Since this example is used to illustrate the consistency between generated reasoning and the predicted trajectory, the evidence is insufficient. In addition, the main experiments compare no reasoning, concise reasoning, and detailed reasoning, but do not report an independent generic-reasoning baseline. The representative outputs for detailed reasoning are also about three times as long as those for concise reasoning and nine times as long as those for the no-reasoning version. The current experiments cannot clearly distinguish gains from the decision semantics provided by the planner from gains due to the additional training signal in longer text. Furthermore, the paper does not sufficiently validate the necessity of the four-slot schema itself through structural ablations of Scene, Constraint, Candidate, and Decision. Overall, the results support the claim that reasoning supervision can improve trajectory prediction, but they do not fully establish the independent contribution of the proposed planner-grounded reasoning in terms of faithfulness and design.

  2. Limited generalization evidence and incomplete experimental validation. All main experiments use Qwen3.5-4B, with no results on other backbones, limiting the assessment of cross-model generalization. The paper also does not report means, variances, or confidence intervals across multiple random seeds, so the stability of some smaller gains remains unclear. In addition, the evaluation mainly focuses on open-loop errors between predicted trajectories and the target trajectories generated by rule-based planners. Although Section 4.4 reports the closed-loop performance of the rule-based planners, it does not evaluate the trained VLA in closed-loop driving. Therefore, the current results do not directly support claims of improved closed-loop safety or interaction capabilities. The planner-source ablation also contains results that are not fully explained: when IDM MOBIL Planner is used as the sole teacher, ADE@3s reaches 3.60 and 3.70 under the three-camera and eight-camera settings, respectively, far higher than in the other settings. The authors attribute this to the planner's limited suitability for certain scenarios, but the target trajectories used for evaluation and the comparability of results across teacher settings remain unclear. Furthermore, the paper does not clearly explain the relationship between filling in missing teacher scores and the actual teacher selection process, or how teacher selection scores relate to the training, validation, and test splits. Together, these issues limit the assessment of result stability, the effectiveness of the teacher selection strategy, and the generalization of the overall method.

Questions for authors
  1. Reasoning faithfulness and supervision source: Beyond the fact that the teacher's execution state and target trajectory come from the same planner execution, have the authors evaluated the consistency between the reasoning generated by the trained model and its actual planning decisions? What are the results for the generic-reasoning baseline? Could the authors clarify whether this experiment was conducted and explain how they distinguish the contribution of planner-grounded reasoning from that of general reasoning supervision and differences in text length?

  2. Generalization and driving performance: Have the authors evaluated the method with other backbones and assessed result stability across different random seeds? In addition, has the trained VLA been evaluated in closed-loop simulation? If not, to what extent can the improvements in open-loop trajectory errors support claims about actual driving decisions and safety?

  3. Planner-source ablation and teacher selection: Can the unusually high errors of IDM MOBIL Planner in Table 2 be further explained? Do experiments using different teacher sources use the same target trajectories for evaluation?

Reviewer_cNkbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Neuro-Symbolic Drive, a framework that trains a vision-language-action (VLA) model using rule-grounded reasoning traces extracted from classical rule-based planners. Unlike post-hoc driving rationales, these traces are derived from the same planner computations that determine the target trajectories, grounding the supervision in the underlying decision process. The framework records traces from three complementary planners and converts them into a unified representation of scene context, active constraints, candidate evaluations, and final decisions which then the model is fine-tuned to generate this structured reasoning followed by the corresponding trajectory. On a simulator-generated benchmark, detailed reasoning supervision improves trajectory prediction over trajectory-only training, reducing ADE at three seconds from 0.47 to 0.26 and miss rate from 8.30% to 6.40% under three-camera perception.

Critical strengths
  1. Neuro-Symbolic Drive extracts reasoning traces and trajectories from the same rule-based planner execution, grounding each training explanation in the decision process that produced its target motion. Its unified representation of scene context, active constraints, candidate evaluations and final decisions makes reasoning from three complementary planners usable as consistent supervision for a VLA. This provides a principled alternative to post-hoc rationales, whose connection to the underlying planning process may be unclear.
  2. Experiments compare trajectory-only, concise-reasoning, and detailed-reasoning supervision across both three-camera and eight-camera settings, providing controlled evidence that planner-derived reasoning improves trajectory prediction. Detailed reasoning reduces three-second ADE from 0.47 to 0.26 with three cameras and from 0.54 to 0.26 with eight cameras, while also lowering miss rates. These ablations strengthen the empirical contribution by examining both the value of intermediate decision structure and the additional benefit of more detailed explanations.
Critical weaknesses
  1. The experiments show that training with planner traces improves the VLA’s trajectory predictions over trajectory-only training. What remains unclear is whether, when making a prediction, the trained VLA actually uses the reasoning it generates to choose its trajectory. It could learn useful driving patterns from the traces during training, then generate a plausible explanation and a trajectory from the same scene information without the explanation affecting the choice. Changing or removing reasoning steps at inference time and checking whether the predicted trajectory changes consistently would help test this.
  2. The evaluation measures trajectory prediction against planner-generated targets but does not assess the trained VLA in closed-loop driving. Although the paper reports closed-loop results for the rule-based teachers, these do not establish the learned model’s performance when its own actions affect subsequent observations. Consequently, lower displacement error and miss rate do not demonstrate improved traffic interaction, recovery from mistakes, or driving safety. Closed-loop evaluation of the trained VLA would be needed to establish whether the prediction gains translate into better driving behavior.
Questions for authors
  1. The planner-derived traces are grounded in the computation that produced the supervision trajectories, but how do you establish that the trained VLA’s generated reasoning causally influences its trajectory predictions?
  2. Have you evaluated interventions that modify, replace, or remove reasoning steps at inference time and measure whether the resulting motion changes consistently?
  3. The authors are encouraged to evaluate the proposed method in closed-loop setting, say Bench2Drive or Fail2Drive, during rebuttal period.
Reviewer_8vA7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Neuro-Symbolic Drive, a method for training driving VLAs using execution traces from rule-based planners. It converts traces from multiple planners into a common reasoning format and uses them to fine-tune a VLA to predict reasoning and future trajectories. Experiments on a simulator-generated dataset show improvements in trajectory prediction over trajectory-only training.

Critical strengths
  • The mismatch between a driving model's explanation and its predicted behavior is a relevant problem. Using rule-based planner execution traces is a sensible approach to addressing this.
  • The paper brings multiple planners into a common reasoning framework and shows substantial improvements in trajectory prediction over training without reasoning.
Critical weaknesses
  • The evaluation primarily measures how closely the VLA matches trajectories generated by the rule-based planners, rather than its overall driving performance. Although the teacher planners are evaluated using closed-loop metrics, the final VLA is not. It is therefore unclear whether the improvements in trajectory prediction translate into better driving behavior.
  • The analysis does not sufficiently establish whether the generated reasoning is actually useful for planning. For example, would changing intermediate reasoning tokens about a traffic light from green to red, while keeping the scene fixed, affect the predicted trajectory? The results also lack multiple training seeds or confidence intervals to establish whether the reported improvements, particularly from concise to detailed reasoning, are consistent.
Questions for authors
  1. What motivates the introduction of CCLS-R instead of using the official nuPlan closed-loop score? Could the authors compare teacher selection using CCLS-R, the official nuPlan score, and perhaps an equal-weight score to show how sensitive the results are to the chosen metric?
  2. Figure 4 would benefit from a more quantitative explanation of the differences between planner trajectories. It is difficult to determine why one planner should be preferred from the visualizations alone.
  3. Could Figure 5 be grouped by maneuver or scenario type? The shuffled scenario indices make it difficult to interpret without repeatedly referring to the appendix. Also, Appendix A.2 mentions filling missing teacher scores with random low values. Could the authors clarify whether these values affected teacher selection or only the visualization?
  4. Could metrics such as ADE and FDE be expanded and defined when first introduced? The miss-rate threshold should also be stated explicitly.
  5. The sample trajectory in Appendix A.4 appears inconsistent with the reported ego-state history. Despite the low initial velocity, the first future waypoint is far from the current position. Could the authors clarify whether the coordinates are expressed in different frames or whether there is an issue with the example?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.