acceptodds
Under review as a conference paper at ICLR 2027

REARL: REFERENCE-GUIDED ONLINE CALIBRATION OF TRAFFIC SIMULATION FOR AUTONOMOUS DRIVING

Abstract

Accurate traffic simulation is important for developing and evaluating autonomous driving systems, but local driving rules and fixed initial conditions can drift from real traffic as a rollout evolves. Existing rule-based, data-driven, and initialization- focused methods either provide limited adaptation during a rollout or replace the native simulator with a learned policy. We propose REARL, a reference-guided online calibration framework that corrects this drift while retaining the simulator’s native controller. REARL builds a library of representative scenes from the highD highway trajectory dataset, retrieves a reference from the observed simulation prefix, and measures speed-distribution and spacing discrepancies at scheduled checkpoints. When a discrepancy exceeds a calibrated threshold, a large language model (LLM) selects one simulator-native action from a time-aligned real-traffic snapshot; otherwise, the Intelligent Driver Model (IDM) with the Minimizing Overall Braking Induced by Lane changes (MOBIL) controller (IDM+MOBIL) continues unchanged. In a paired development study on highD scenes, REARL– Qwen reduces speed Hellinger distance from 0.3313 to 0.2945 and spacing error from 0.9190 to 0.8448 relative to Base. Component controls show that updating the reference improves speed alignment, deterministic correction achieves lower spacing error at a substantial jerk and headway cost, and a retrospective matched- budget periodic comparator has lower error than sparse discrepancy triggering. These results support online reference-guided calibration as a practical middle ground between fixed initialization and full policy replacement.

Primary area:
applications to robotics, autonomy, planning
Keywords:
traffic simulation, closed-loop calibration, autonomous driving, large language models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_e8uMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes REARL, a reference-guided online calibration framework for improving the realism of traffic simulation. The method constructs a representative scene library from real-world traffic data and dynamically retrieves reference scenes based on the current simulation state. By detecting discrepancies in speed distributions and vehicle spacing, REARL employs an LLM to select corrective actions while preserving the simulator's original IDM+MOBIL controller. Experiments on the highD dataset demonstrate that REARL improves the statistical alignment between simulated and real traffic, particularly in terms of speed distribution and vehicle spacing.

Critical strengths
  1. Modular design. The proposed framework preserves the simulator's native controller and introduces a separate online calibration module, allowing integration without retraining the underlying driving policy.
  2. Detailed component analysis. The paper includes several ablation studies and diagnostic metrics to examine the effects and limitations of different components.
Critical weaknesses
  1. High computational overhead. Although the paper reports an efficiency comparison, REARL introduces substantial computational overhead due to repeated LLM inference for individual vehicles. This raises concerns about its practicality and scalability for large-scale traffic simulation and autonomous driving evaluation.
  2. Questionable effectiveness of adaptive interventions. The full REARL configuration triggers interventions at every scheduled checkpoint, effectively resulting in periodic corrections. Moreover, the ablation study shows that periodic intervention can outperform discrepancy-triggered intervention under a matched budget. These results raise questions about the necessity and effectiveness of the proposed adaptive intervention mechanism.
  3. Lack of downstream validation. Although REARL improves statistical alignment with real-world traffic, it remains unclear whether these improvements benefit the training or evaluation of autonomous driving models. Experiments involving downstream autonomous driving policies or closed-loop evaluation would better demonstrate the practical value of the proposed framework.
  4. Limited definition of traffic realism. The evaluation primarily relies on speed distributions and vehicle spacing to measure simulation realism. However, matching these statistical properties does not necessarily indicate realistic vehicle interactions or driving behaviors. More comprehensive evaluations of behavioral realism, particularly in safety-critical interactions, are needed.
  5. Limited evaluation scenarios and generalization. The experiments are conducted on only five traffic scenes. Moreover, these scenes were also used during prompt development, raising concerns about potential overfitting. Evaluation on a larger and more diverse set of unseen traffic scenarios is necessary to demonstrate the generalization and robustness of the proposed framework.
Questions for authors

None.

1 comment

Cost. Repeated vehicle-level calls are expensive. We did not take a new wall-clock measurement. The figure we cite is the ten-step benchmark, 86,241.309 ms against 0.461 ms for IDM+MOBIL and 15.898 ms for PPO, and it applies to offline calibration. Full-setting rollouts in the new runs use 110 calls each. We do not claim that this cost is appropriate for large-scale traffic simulation or online autonomous-driving evaluation.

Adaptive timing. The full configuration fires through the spacing branch at every scheduled checkpoint, so the realized schedule is periodic. That is already stated in the paper. Both the retrospective matched period and the prospective DeepSeek comparison give lower corresponding errors to the periodic side. An advantage for adaptive timing is not established. The speed and spacing improvement over IDM+MOBIL remains, including on the ten new scenes.

Downstream driving. No downstream policy was trained or evaluated in closed loop. Whether the statistical alignment helps autonomous-driving training or evaluation is unanswered. We do not report a downstream benefit.

Realism. The reported quantities are the speed distribution, Equation (2), the same-lane gap distribution, jerk-distribution distance, and the colliding-vehicle rate. Across the five development scenes, mean absolute jerk of the real trajectories lies between 0.059 and 0.071. One full-bin acceleration or deceleration moves the target speed by 3.375 m/s. Agreement on these statistics does not establish realistic interaction, and this revision adds no evaluation on scenes built for safety-critical events.

Coverage. The submitted evaluation used five scenes that also informed prompt design. The ten additional scenes did not. Their Qwen, DeepSeek, and GLM results are separated by interface in Section 1. The further 15 reserved scenes were not run, so the added evidence stops at those ten scenes.

Reviewer_qHhqOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a reference-guided correction layer for traffic simulation. It retrieves a highD scene using the observed simulation prefix, compares speed and spacing statistics, and uses an LLM to select brief simulator-native actions while otherwise retaining IDM+MOBIL. The evaluation uses five development scenes with five simulation seeds each. The strongest supported finding is that updating the reference improves speed-distribution alignment within this development suite.

Critical strengths

1、The distinction between a changing online reference and a fixed evaluation reference is well motivated. It prevents apparent improvements from being produced merely by changing the target against which a rollout is scored (Section 3.1). 2、The component study is unusually candid. It reports the deterministic controller's spacing-smoothness trade-off, the retrospective nature of the matched-budget timing control, and the absence of evidence that adaptive triggering is beneficial (Table 2 and Appendix F). 3、The reported Qwen configuration improves speed Hellinger distance from 0.3313 to 0.2945, with a speed improvement in all five scene means. Additional gap, jerk, collision, and temporal diagnostics help interpret this result (Tables 1 and 5; Figure 2).

Critical weaknesses
  1. The spacing objective does not measure agreement with the reference in a straightforward way. Equation 2 compares every reference vehicle's mean pairwise distance with one scalar: the simulated scene's overall mean distance. Even exact replay, x = y, generally has positive error because the reference vehicles have different mean distances. For an unconstrained scalar summary, the numerator is minimized at the median of the reference distances, rather than necessarily at the summary produced by replay. Distinct spatial configurations can also receive identical scores. This does not invalidate the speed results, but it makes a reduction in the primary spacing objective difficult to interpret as improved traffic fidelity. The same objective also drives retrieval and triggering. The additional same-lane-gap diagnostic is useful, but does not resolve this definition issue.
  2. There is no independent evaluation of the final configuration. Sections 4.1 and B explicitly state that the five evaluated scenes informed prompt selection. The 25 rollouts therefore provide five development-scene units, not 25 independent scene tests. The scene-level bootstrap is appropriate for the reported within-suite comparison, but cannot remove selection bias. Recording-disjoint preprocessing alone does not establish generalization of the final prompts and controller.
  3. The incremental value of the proposed correction mechanism remains insufficiently isolated. All 275 full-setting checkpoints trigger through the spacing branch, and the periodic matched-budget control outperforms sparse triggering. The deterministic non-LLM policy improves spacing more, although with worse jerk. A smoothness-constrained deterministic correction is a particularly relevant missing control for the fidelity-cost trade-off. PPO and CRITICAL use different simulation semantics, so their main-table results cannot establish a controlled advantage. These limitations matter given the system-level timing of 86,241 ms reported in Appendix E; low latency is not required, but the additional cost should buy a demonstrated benefit.
Questions for authors
  1. Can you report Equation 2 on exact replay and explain its intended optimum? Does the method ranking persist with a distributional spacing metric that is zero for identical distributions, using the existing rollout records?
  2. Is there a set of scenes untouched by prompt or controller development on which the final configuration can be evaluated without retuning? Please distinguish these results from the development results.
  3. At a fixed intervention budget, how does REARL compare with a deterministic controller constrained to comparable action magnitudes or smoothness? Which result specifically supports using the LLM?
1 comment

Q1. Exact replay has mean Equation (2) error 0.1896, and the median scalar averages 0.1819. We keep Equation (2). On mean relative error and vehicle-distance W1, both zero when the quantities they compare agree, the five-scene order remains rule, Qwen, DeepSeek, Base, GLM. The speed order differs. A lower S is not reported as greater realism. The GLM row in this comparison is the November extract, not Table 1 and not the scaled-step run.

Q2. This is the same question as YBQZ Q1. The ten new scenes were not used to develop the prompt or the controller, and they were evaluated without retuning. The fresh development-scene calls are reported separately. The reserved 15 scenes have no result.

Q3. A deterministic controller constrained to comparable action magnitude or smoothness, at a matched intervention budget, was not run. The published deterministic rule has lower spacing error and jerk W1 0.577, against 0.189 for Qwen. The result that supports the correction layer is the reduction in speed and spacing error relative to uncorrected IDM+MOBIL, including on ten scenes that did not inform the prompts. A matched experiment that would show an LLM is required is not available.

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

REARL builds a library of representative highD scenes, retrieves a reference using the observed simulation prefix, and compares speed and spacing statistics at scheduled checkpoints. When a threshold is exceeded, an LLM chooses a discrete simulator-native correction separately for ten vehicles; IDM+MOBIL resumes afterward. Evaluation uses a fixed real-scene target distinct from the changing online reference, with five development scenes crossed with five simulation seeds and several component substitutions.

Critical strengths
  1. Separating the online reference from the fixed evaluation target is an important design choice: a method cannot improve its score merely by choosing an easier reference for itself. The paired comparison with the unchanged native controller supports a real effect within this setup. Qwen reduces speed Hellinger distance from 0.3313 to 0.2945 and spacing error from 0.9190 to 0.8448 (Table 1). Updating the reference also improves scene-mean speed alignment over freezing it in all five scenes.
  2. The component study reports meaningful negative results and behavioral costs. The deterministic correction achieves lower spacing error but substantially higher jerk, while the retrospective periodic comparator outperforms sparse discrepancy triggering at matched intervention counts. The paper explicitly acknowledges prompt-development reuse, unmatched PPO/CRITICAL simulation semantics, and the absence of a safety guarantee. Scene-level bootstrap resampling and the paired temporal curves are more informative than presenting 25 runs as 25 independent environments.
Critical weaknesses
  1. There is no independent evaluation of the selected system. The five reported scenes subsequently informed prompt design, even though their recordings were excluded from clustering and threshold calibration. Five simulation seeds per scene do not remove that selection dependence, and bootstrap intervals over the same five scenes cannot establish generalization. All paired evidence concerns one four-lane, ten-vehicle, 60-step setup; it does not establish reliable calibration across the available scene pool. The disclosure is appropriate, but a development-only result is insufficient support for the practical scope of the proposed framework.
  2. The distinctive mechanisms and cost are not justified by the present controls. Every full-setting checkpoint triggers through spacing, so the main result does not test selective intervention or the speed-trigger branch. The retrospective periodic result appropriately does not establish a deployable alternative, but neither does it support adaptive timing. The single deterministic comparator is not matched for conservative behavior or a jerk constraint, leaving the value of LLM action selection unresolved. Moreover, Eq. (2) compares real per-vehicle mean distances with one simulated mean-distance statistic; different spatial configurations can share that statistic. The improved gap diagnostic helps, but jerk worsens from 0.138 to 0.189 and collision rates remain similar. With a reported system timing of 86,241 ms in the ten-step benchmark, evidence of an advantageous fidelity-cost trade-off is still needed even for offline simulation calibration.
Questions for authors
  1. Can the frozen library, prompts, parser, and thresholds be evaluated on scenes not used for prompt development, using scene-level paired results and the same simulation semantics? Please distinguish an actual untouched holdout from another evaluation on the development scenes.
  2. Does the full configuration outperform a prospectively specified periodic correction policy with the same call budget and a conservative, smoothness-constrained non-LLM action rule? A matched snapshot-removal or shuffled-reference control would clarify whether the reference content, rather than the conservative correction prompt alone, explains the benefit.
  3. Can the authors report the calibration-cost trade-off on the independent cases, including actual action calls, system runtime, gap and jerk discrepancies, and collision statistics? The decision-relevant question is whether the narrower reference-guided correction mechanism improves useful traffic behavior at a defensible cost.
1 comment

Q1. The ten scenes drawn with seed 20261111 did not inform prompt design. The library, prompts, parser, thresholds, and simulation semantics were held fixed. Their means and scene counts are reported separately from the fresh development-scene calls. The 15-scene mean pools the two groups, so it is not itself a pure holdout. The further 15 reserved scenes were not evaluated. Bootstrap intervals on the original five scenes still describe those five scenes.

Q2. A periodic policy with the same call count as the full configuration was finished for 15 of 25 Qwen rollouts. Both sides use 11 interventions, and the score gap is another sample. On DeepSeek, the six-checkpoint period fixed in advance is lower on all five means than sparse P60, at 60 calls versus 53.6. The completed prospective comparisons do not support discrepancy triggering. A smoothness-constrained non-LLM rule, snapshot removal, and a shuffled reference were not run. The published freeze-reference control remains: Qwen H rises from 0.2945 to 0.3147 and S from 0.8448 to 0.8753 when the first retrieved scene is kept. That control stops reference updates. It does not shuffle reference content.

Q3. On the ten new scenes, Qwen, conservative DeepSeek, and scaled-step GLM each use 110 calls and 11 interventions per rollout. No new wall-clock measurement was taken. The submitted ten-step benchmark remains 86,241.309 ms for REARL, 0.461 ms for IDM+MOBIL, and 15.898 ms for PPO. It is an offline calibration cost and includes retrieval, discrepancy computation, and API reasoning. On those ten scenes Qwen lowers gap Hellinger from 0.5011 to 0.4226 and mean collision rate from 0.2200 to 0.1680, while mean jerk rises from 0.1418 to 0.1604. The stable gain at this cost is closer speed and spacing, not uniformly better traffic behavior. The review cites Base jerk as 0.128. Table 1 lists Base jerk as 0.138 and Qwen jerk as 0.189.

Comments on the paper

1 comment

We ask the Area Chair to judge the paper on its actual claim. A short correction, guided by a real scene retrieved from the observed prefix and then returned to IDM+MOBIL, brings speed and spacing closer to a fixed real evaluation scene than the uncorrected controller. That scene is fixed: it is not given to the controller, and it is not the retrieved scene. The revision tests the frozen library, prompts, parser, and thresholds on scenes that never entered prompt design.

The cleanest new evidence is Qwen under the submitted interface, with all fifteen scenes rerun. On the ten scenes held out from prompt design: IDM+MOBIL has H 0.3351, S 0.9248, Gap H 0.5011, Jerk W1 0.1418, and collision rate 0.2200; Qwen has 0.3093, 0.8478, 0.4226, 0.1604, and 0.1680. Speed is closer in 10/10 scenes, spacing in 9/10, and same-lane gap distribution in 10/10. Over fifteen scenes the same direction holds for speed (15/15), spacing (14/15), and gap distribution (15/15). Mean collision is lower and mean jerk is higher; we report that cost. DeepSeek, under the same conservative instruction, is also below IDM+MOBIL, BC-GRU, HAD-Gen, and PPO on speed, spacing, gap distribution, jerk, and collision, with the lowest jerk in that comparison, 0.1262. These figures supplement Table 1.

Two boundaries were already in the paper. We do not claim that discrepancy-triggered timing beats a fixed period. The published matched-budget control already favors the period, and a prospective DeepSeek comparison, specified before either arm and finished for all 25 rollouts, agrees: six fixed checkpoints beat sparse P60 triggering on all five means. We also do not claim that the corrector must be an LLM. The published rule lowers spacing error from 0.8448 to 0.6486 and raises jerk W1 from 0.189 to 0.577. The LLM is the better smoothness trade-off, not proof that language reasoning is required.

We agree with the mathematics of Equation (2). Exact replay has mean error 0.1896, because one simulated scalar is compared with every real vehicle mean distance; the median target averages 0.1819, and different layouts can share the scalar. Speed uses a separate distributional distance, so this objection does not reach it. On the submitted rollouts, two spacing measures that are zero when their inputs agree keep the five-scene order of Equation (2): rule, Qwen, DeepSeek, IDM+MOBIL, then the November GLM extract. We kept the prespecified metric. A lower value is a smaller prespecified error, not proof of realistic behavior in every sense.

The other requests sit outside this claim. A smoothness-limited non-LLM rule would be a new controller. The paper already removes the LLM and reports both the further spacing gain and the jerk cost. Our claim is the short correction around IDM+MOBIL, which holds on the original scenes and on the ten new ones. Removing the snapshot or shuffling the reference would test incidental prompt text. No correction returns the rollout to IDM+MOBIL, and freezing the first retrieved scene raises Qwen speed error from 0.2945 to 0.3147, so the online choice of reference already matters. The reserved fifteen scenes would repeat the same four-lane, ten-vehicle protocol; the ten we ran were fixed in advance and were not used to write the prompts. Ten Qwen periodic calls returned HTTP 403. On the fifteen that finished, both the period and the full configuration intervene at all eleven checkpoints, so more seeds cannot show that the trigger chooses a different time. The finished DeepSeek run is the timing evidence. Downstream driving and safety-critical suites evaluate a use we do not claim: REARL calibrates ordinary HighD background traffic and trains no planner. Jerk rises for submitted Qwen, and collision rate is not a safety guarantee. We did not retime the new scenes. Each full-setting rollout makes the same 110 calls, and the submitted benchmark already includes retrieval, discrepancy, and API reasoning: 86,241.309 ms, against 0.461 ms for IDM+MOBIL and 15.898 ms for PPO, as an offline-calibration cost.

The reviews ask for superior adaptive timing, a necessary LLM, a downstream-driving gain, and a safety proof. The paper draws none of these, and the revision adds none. On the claim it does make, reference-guided correction moves speed and spacing closer to the fixed real scene than uncorrected IDM+MOBIL, the ten new scenes agree, and the stated limits remain. We ask the Area Chair to weigh the paper on that ground.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.