acceptodds
Under review as a conference paper at ICLR 2027

Learning Nonlinguistic Deliberation Through Latent Critiques

Abstract

Planners often use a learned scorer to select among candidate actions, then discard its assessment. We ask whether this assessment can instead guide revision, and what the critic should communicate to make revision useful. We introduce ReCoursive, a recurrent planner for end-to-end driving that deliberates without language. Its critic returns latent critiques, the continuous representations from which it predicts scores, to guide the proposer’s subsequent revisions. The loop is trained end to end with trajectory and assessment supervision at each step, without reasoning traces or demonstrated corrections. On NAVSIMv2, critic feedback adds 2.83 EPDMS over recurrence alone at matched depth and proposal count. Additional revision steps improve planning with latent critiques but not with predicted scores alone; critiques also need assessment supervision and must be refreshed as candidates change. ReCoursive achieves the best results among compared methods on NAVSIMv2 and, in zero-shot transfer, on HUGSIM and among non-VLM methods on KITScenes LongTail. These results suggest that nonlinguistic deliberation benefits from a learned interface between assessment and generation rather than one restricted to scores.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Nonlinguistic Deliberation, Latent Feedback, Iterative Refinement, End-to-End Autonomous Driving, Test-Time Computation

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

The paper introduces ReCoursive, a recurrent end-to-end driving planner that performs nonlinguistic deliberation through latent critiques. A weight-shared proposer generates candidate trajectories, and a critic both scores them and returns continuous per-candidate representations. These latent critiques condition the next proposal step, creating a propose–assess–revise loop. The model is trained end to end with trajectory and assessment supervision at every deliberation step, without language reasoning traces or demonstrated corrections. At inference, metric heads rank all cumulative candidates for final selection.

Critical strengths

The paper reframes scorer-based planning as a propose–assess–revise loop and identifies the assessment-to-revision interface as the key design choice. Returning the critic’s latent representation, rather than only scalar scores, is intuitive and well aligned with the goal of nonlinguistic deliberation.

Critical weaknesses

The paper exhibits extensive traces of AI assistance.

The assessment-supervision ablation uses a common frozen no-feedback critic, but the gap already appears at depth one, before later feedback-driven revision can explain it. Removing metric losses may also change the shared scene representation during training, so the depth-six gain cannot be attributed solely to better critiques. The explicit-score feedback comparison is also ambiguous: explicit-score feedback starts higher at depth one but fails to improve, which may reflect optimization difficulty rather than a fundamental limitation of scalar feedback. Finally, the feedback-editing interventions are illustrative examples, and the authors themselves state that they do not establish reliable directional control.

Questions for authors

Missing recent SOTA baselines. The comparison tables omit several recent methods directly comparable on NAVSIM-v2 navhard/navtest and NAVSIM-v1 navtest. Could the authors clarify why the following are not included? On navhard: DriveFuture (55.5 EPDMS, rank 1 as of April 2026), GraphWorld (53.6 EPDMS), HiST-VLA (50.9 EPDMS), MomWorld, MomADv2 (39.5 EPDMS), PhysWAM (38.1/39.8 EPDMS), GeoWAM (36.6 EPDMS), 4D-WAM (35.9 EPDMS), PerceptDrive (34.5 EPDMS), RAP (39.6 EPDMS), IDOL (76.2 EPDMS stage-1). On navtest: ReDrive (90.8 EPDMS), 4D-WAM (90.6 EPDMS), GeoWAM (90.2 EPDMS), PerceptDrive (90.2 EPDMS), ManeuverFlow (90.1 EPDMS), ForeDrive (90.0 EPDMS), SparseDriveV2 (90.1 EPDMS), HiST-VLA (88.6 EPDMS), Gaze-Flow (86.7 EPDMS), DriveSuprim (87.1 EPDMS), DriveWorld-VLA (86.8 EPDMS), IDOL (89.6 EPDMS), WA-JEPA (91.7 EPDMS). On NAVSIM-v1: TOAD (94.7 PDMS), DriveSuprim (93.5 PDMS), RAP-DINO (93.8 PDMS), SparseDriveV2 (92.0 PDMS), iPad (91.7 PDMS), Gaze-Flow (91.5 PDMS), DriveFuture (90.7 PDMS), 4D-WAM (90.9 PDMS), PerceptDrive (90.4 PDMS), ManeuverFlow (90.6 PDMS), ForeDrive (89.9 PDMS), GraphWorld (90.1 PDMS), ReDrive (91.0 PDMS), DriveWorld-VLA (91.3 PDMS). On HUGSIM: WA-JEPA (HD-Score 0.4462), PhysWAM (RC 48.9 / HD-Score 35.5), TOAD.

3 comments

Part 1/2

We thank reviewer TfvF for the constructive review and the effort invested. We are glad to address the concerns raised.

The paper exhibits extensive traces of AI assistance.

We used AI tools as described in our AI Use Statement. We reviewed all AI-assisted content and revised for a natural academic style throughout. Choices such as opening each paragraph with a bold topic sentence were deliberate to help human readers skim the paper easily.

The assessment-supervision ablation uses a common frozen no-feedback critic, but the gap already appears at depth one, before later feedback-driven revision can explain it. Removing metric losses may also change the shared scene representation during training, so the depth-six gain cannot be attributed solely to better critiques.

We agree that removing metric losses might change the learned scene representation (Appendix D.2). Our results show that assessment-supervised training is associated with larger gains across recurrent revisions, but do not fully isolate the semantic contribution of assessment supervision.

Complementarily, to isolate the role of downstream planning gradients, we performed a stop-gradient ablation that retains the forward feedback interface and all assessment losses. This substantially reduces both final performance and the improvement obtained across revisions, supporting the importance of training the critic through its use in the proposer. Please see the ablation results below.

The explicit-score feedback comparison is also ambiguous: explicit-score feedback starts higher at depth one but fails to improve, which may reflect optimization difficulty rather than a fundamental limitation of scalar feedback.

We agree that the feedback interface affects the optimization difficulty. To test this, we compared three explicit-score interfaces: logit input, an MLP projection, and the paper's affine projection. In addition, in the stop-gradient experiment mentioned above, we cut the gradient from trajectory losses to the critic. In this way, the critic is only supervised by the assessment losses, without being optimized by the trajectory losses. We found that the stop-gradient model performs similarly to the no-feedback model, indicating that effective revisions rely on the critic being trained by trajectory losses to provide guidance beyond explicit assessments.

All results below are single-seed cross-fit EPDMS on navhard (mean [min, max] over 3 fold seeds, following the protocol in the common response above):

Variantdepth 1depth 2depth 6depth 6 - depth 1
Explicit scores (paper, affine projection)47.26 [46.30, 48.25]47.27 [46.64, 47.95]46.79 [45.79, 47.45]-0.47 [-0.80, -0.09]
Explicit scores, logit input45.69 [45.25, 46.25]45.95 [45.49, 46.23]46.99 [46.68, 47.53]+1.31 [+1.13, +1.52]
Explicit scores, MLP projection49.43 [49.24, 49.80]50.79 [50.65, 50.97]50.11 [49.99, 50.20]+0.68 [+0.33, +0.97]
Stop-gradient latent feedback46.03 [45.41, 46.77]46.87 [46.76, 47.00]47.05 [46.03, 47.97]+1.02 [+0.62, +1.23]
No feedback45.87 [45.04, 46.89]47.09 [46.67, 47.69]47.12 [46.55, 47.47]+1.25 [+0.57, +1.67]
Latent feedback (ReCoursive-S)48.98 [48.72, 49.15]52.49 [52.43, 52.58]52.47 [52.39, 52.61]+3.49 [+3.33, +3.68]

Takeaways:

  1. Stop-gradient performs similarly to no feedback and substantially underperforms ReCoursive-S. The result aligns with our claim that the critic must learn from trajectory losses to guide revisions.
  2. Explicit-score variants yield at most the recurrence gain of no feedback.
  3. ReCoursive-S achieves both higher final EPDMS and substantially larger revision gains than all explicit-score variants.

Note on evaluation convention: At bzZK and QBm8's request, Sections 4.2–4.4 now report 5-fold cross-fit EPDMS, while Table 2 keeps the standard navhard full-set fitting for direct comparability with published NAVSIM works (with cross-fit alongside in Table 12). Given your familiarity with the NAVSIM literature, we welcome your view on keeping both protocols.

Part 2/2

Finally, the feedback-editing interventions are illustrative examples, and the authors themselves state that they do not establish reliable directional control.

We include the feedback-editing interventions to show that the feedback affects revisions, not for introducing a method to control the revision reliably. To quantify this beyond single scenes, we sampled 66 scenes where the ego vehicle is close to the drivable area boundary for the DAC editing, and 85 scenes with surrounding vehicles for NoC & TTC editing:

EditMetricScenesSuccess (Safety changed)Trajectory endpoint changed > 0.1 m
Toward unsafeDAC6637 (56%)66 (100%)
Toward safeDAC6634 (52%)66 (100%)
Toward unsafeNoC8552 (61%)85 (100%)
Toward safeNoC8529 (34%)85 (100%)

"Success" means we successfully push a safe candidate to unsafe, or push an unsafe one to safe, using the flipped donors. We note that although many edits do not change the safety label, all of them shift the trajectory endpoint by more than 0.1 m.

Q1. Missing recent SOTA baselines. The comparison tables omit several recent methods directly comparable on NAVSIM-v2 navhard/navtest and NAVSIM-v1 navtest. Could the authors clarify why the following are not included?

Baseline selection criterion: The ICLR 2027 reviewer guidelines treat papers published on or after July 17, 2026 as contemporaneous and do not require comparisons with papers available only on arXiv.

On navhard, we went further in two ways:

  1. Verifiable prior baselines: We included methods that, by July 17, 2026, were either peer-reviewed or had released code and weights for navhard (LTF, LTFv6, RAP, ZTRS, DrivoR, SimScale), enabling reproducible comparisons.
  2. Same-backbone extensions (regardless of cutoff date or code release): We also included the two recent works that directly build on our DrivoR base planner: TOAD (which refines proposals with DrivoR's scorer via CEM search) and Gigapixel-DrivoR (which adapts DrivoR through self-play rather than imitation learning).

NAVSIM-v2 vs. NAVSIM-v1: We selected benchmarks to cover 3 complementary evaluation settings (see line 299): pseudo closed-loop (NAVSIMv2), closed-loop (HUGSIM), and longtail scenarios (KITScenes-Longtail). NAVSIMv1 overlaps with NAVSIMv2 on the same nuPlan domain, while NAVSIMv2 upgrades single-stage PDMS to two-stage pseudo-simulation and richer metrics (EPDMS), so we opted for NAVSIMv2.

navhard vs. navtest: navhard is more challenging and less saturated (SOTA methods achieve over 90 on navtest, but below 60 on navhard). Therefore we chose navhard.

For the works you suggested on navhard and HUGSIM, we applied the same criterion:

  • Already included: RAP (Table 2) and TOAD (Tables 2 and 3).
  • Appeared after July 17, 2026: PerceptDrive, 4D-WAM, WA-JEPA, MomADv2, GeoWAM, PhysWAM and MomWorld. PhysWAM and MomWorld appeared on arXiv after the ICLR 2027 submission deadline.
  • Appeared before July 17, 2026, but neither peer-reviewed nor with released code or weights: DriveFuture, HiST-VLA, GraphWorld and IDOL.

Revisions: We have added this criterion to Appendix C.7. Although we do not compare against these works in our tables, we have cited all of them in the Related Work of the revised paper. This follows the ICLR reviewer guidelines, which do not require comparisons with contemporaneous work or papers available only on arXiv, but strongly encourage authors to cite such related work.

Dear Reviewer TfvF, thank you again for your feedback. With the discussion period closing in about 24 hours, we wanted to check whether you have had a chance to review our responses above and the revised manuscript. Please let us know if any questions remain while we still have time to reply.

Reviewer_QBm8OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces ReCoursive, a recurrent autonomous driving planner that extends trajectory generation and scoring with assessment-conditioned revision. Instead of using a learned scorer solely to select candidates, it feeds the critic’s pre-head latent representations back into a weight-shared proposer to generate subsequent candidates. The unrolled model is trained with trajectory imitation and metric supervision at each iteration, without language reasoning traces or demonstrated corrections. Final selection considers candidates accumulated across iterations.

Validation includes NAVSIMv2, zero-shot closed-loop evaluation on HUGSIM, and zero-shot evaluation on KITScenes LongTail. Matched recurrent controls and feedback ablations investigate the contributions of latent feedback, assessment supervision, and refreshed critiques. The results support the usefulness of this feedback interface within the evaluated architecture, although the strength of the broader claims about deliberation and test-time scaling remains uncertain.

Critical strengths
  1. A focused design question supported by targeted ablations. The paper studies whether a scorer’s internal assessment representations can improve subsequent trajectory generation, rather than merely rank existing candidates. The matched no-feedback recurrent control preserves iteration depth and proposal count, providing meaningful evidence that the gains are not explained solely by additional recurrence or more candidates. The explicit-score and stale-feedback comparisons further strengthen this empirical contribution.

  2. A practical implementation with transparent reporting. The approach reuses visual features and shares parameters across iterations, providing a relatively inexpensive mechanism for iterative refinement without requiring reasoning traces or correction demonstrations. The HUGSIM results provide useful closed-loop evidence beyond the primary benchmark. The paper is also clearly written and unusually explicit about saturation, proxy-label inaccuracies, and the limited conclusions supported by its intervention experiments, which improves the credibility and assessability of the work.

Critical weaknesses
  1. The evaluation protocol limits confidence in the reported gains and scaling conclusions. Appendix C.5 states that selection coefficients are optimized on all 5,912 navhard evaluation scenes using the reported EPDMS objective. These coefficients directly determine the selected trajectory, so this introduces evaluation-set overfitting risk even without updating network parameters. Applying similar searches to released baselines helps assess comparative fairness, but does not establish held-out generalization. This matters because some competitive margins are modest: approximately one EPDMS point over the reproduced TOAD result, and 0.06 MMS over DMAD on LongTail. Moreover, improvements saturate, the eight-step training run underperforms the six-step run, and inference beyond the trained unrolling depth is not established. The evidence therefore supports useful early refinement more strongly than robust test-time scaling.

  2. The validity of the assessment signal and the distinctiveness of the proposed mechanism remain insufficiently established. Training targets use cached future object occupancies to evaluate alternative ego trajectories. Recorded futures are conditioned on the behavior that actually occurred and may not remain valid when the ego vehicle changes its actions, particularly in yielding and negotiation scenarios. This raises a counterfactual validity concern rather than an objection to future-outcome supervision itself. The supervised objectives also do not explicitly capture several important aspects of driving quality, including yielding, right-of-way reasoning, and richer navigation intent; comfort is excluded from supervision and selection. Consequently, improvements under these objectives do not establish that the feedback reliably guides broader behavioral correction. Given the iterative refinement and self-editing literature discussed in the paper, stronger evidence is also needed that the latent interface provides a distinctive assessment-driven correction mechanism rather than an effective additional feature pathway.

Questions for authors
  1. Evaluation protocol — clarification and decisive validation. Appendix C.5 indicates that selection coefficients are fitted on the same navhard scenes used to report EPDMS. Is there an independent split on which these coefficients are determined? If not, can you report results with coefficients fitted on a disjoint validation split and frozen for held-out evaluation? Whether the main gains survive this separation would materially affect my assessment.

  2. Counterfactual target validity — clarification and supporting evidence. Are future object occupancies held fixed when generating training labels for different candidate ego trajectories? If so, what evidence supports their validity in interactive situations where other agents would respond differently to alternative ego actions? Please distinguish agreement with a fixed-future evaluator from evidence of correctness under action-dependent interactions.

  3. Scope of the behavioral gains — clarification of existing results. Can you provide safety, progress, and comfort breakdowns, together with relevant failure cases, to clarify what drives the aggregate improvements? In particular, do the gains extend to interactive yielding and right-of-way situations that are not explicitly represented in the supervised metric heads?

  4. Mechanism and novelty — targeted validation. Does the benefit persist when later planning gradients are detached at the latent feedback interface, while assessment supervision and forward feedback remain unchanged? This would help establish whether training the critic through its use in revision is consequential, and sharpen the distinction from a supervised feature-feedback refinement architecture.

  5. Scaling and robustness — clarification of the supported claim. Are the main feedback gains and the six-versus-eight-step results reproducible across training seeds? Have you evaluated inference beyond the training depth? Please clarify whether the intended claim is refinement within a trained unrolling horizon or generalizable test-time scaling; evidence for the latter would substantially strengthen the contribution.

5 comments

Part 1/3

We thank reviewer QBm8 for the constructive review and clear validation questions. We are glad to address the concerns raised.

Q1. Evaluation protocol — clarification and decisive validation. Appendix C.5 indicates that selection coefficients are fitted on the same navhard scenes used to report EPDMS. Is there an independent split on which these coefficients are determined? If not, can you report results with coefficients fitted on a disjoint validation split and frozen for held-out evaluation? Whether the main gains survive this separation would materially affect my assessment.

The navhard split does not provide a validation set. We therefore have updated the analysis and ablations in paper with group-disjoint 5-fold cross-fitting. It confirms that both the feedback advantage and our lead over TOAD persist. Please see the common response above for the full comparison table, and Sections 4.2–4.4 of the revised paper, which now report cross-fit EPDMS across all ablations.

Q2.1 Counterfactual target validity — clarification and supporting evidence. Are future object occupancies held fixed when generating training labels for different candidate ego trajectories?

Yes, we use fixed future proxy for training, which is a trade-off between training speed and target accuracy. For details please see Appendix B.3 and B.4.

Q2.2 If so, what evidence supports their validity in interactive situations where other agents would respond differently to alternative ego actions? Please distinguish agreement with a fixed-future evaluator from evidence of correctness under action-dependent interactions.

While the training target relies on a fixed-future proxy for tractability, the actual NAVSIMv2 evaluation is dynamic and reactive (using IDM).

We have added "agreement with reactive (PDM) simulator targets" to Appendix B.3 (Table 8). On the same 500 training scenes and 32,000 fixed trajectories, we scored every trajectory twice with NAVSIMv2 PDM simulation: with background agents replaying the log (fixed future), and with IDM agents that react to the ego, as in navhard evaluation. Binary agreement (%; EP: Spearman ρ\rho):

HeadOurs vs log-replayOurs vs reactiveLog-replay vs reactive
NoC87.5286.4391.56
DAC93.6693.66100.00
DDC83.2383.23100.00
TTC85.7683.7491.18
EP (ρ\rho)0.690.660.82

Our targets agree with reactive targets on 83–94% of binary labels. This is only 1.1 points (NoC) and 2.0 points (TTC) lower than with log-replay targets; DAC and DDC do not depend on other agents. The targets remain a fixed-future approximation: when agents react, the simulator's own NoC and TTC labels change for about 9% of trajectories.

NAVSIMv2 evaluation provides further evidence under action-dependent interaction: Against the no-feedback control, latent feedback gains 4.65 EPDMS in unsignalised right-of-way scenes, 3.92 at crosswalks, and 2.43 at signalised intersections (please see Q3).

This evidence shows that a planner trained on fixed-future targets improves under reactive evaluation. The gains in these three contexts are also smaller than in the remaining scenes (6.70), so fixed-future targets may limit the benefit there.

Part 2/3

Q3. Scope of the behavioral gains — clarification of existing results. Can you provide safety, progress, and comfort breakdowns, together with relevant failure cases, to clarify what drives the aggregate improvements? In particular, do the gains extend to interactive yielding and right-of-way situations that are not explicitly represented in the supervised metric heads?

We report the depth-6 cross-fit submetrics on NAVSIMv2.

No feedbackReCoursive-SΔ [min, max]
No Collision (NoC)91.2292.93+1.71 [+1.21, +2.05]
Drivable Area Compliance (DAC)83.6886.61+2.93 [+1.81, +3.61]
Driving Direction Compliance (DDC)95.0896.49+1.41 [+0.36, +1.97]
Traffic Light Compliance (TLC)98.9899.41+0.42 [+0.17, +0.61]
Time To Collision (TTC)89.6491.30+1.66 [+1.18, +1.99]
Ego Progress (EP)43.7644.45+0.69 [-0.61, +2.71]
Lane Keeping (LK)49.9353.15+3.22 [+1.52, +4.18]
Comfort (Comf)96.3396.22-0.11 [-0.41, +0.14]
Extended Comfort (EC)55.1953.70-1.49 [-2.41, -0.14]
EPDMS47.1252.47+5.34 [+4.92, +5.85]

Takeaways:

  1. ReCoursive-S achieves better subscores in all submetrics except two comfort submetrics.
  2. TLC and LK are not trained or considered in proposal selection, yet both improve across all fold seeds.

We clarify that our method was not specifically designed to address interactive yielding or right-of-way negotiation, nor does NAVSIMv2 isolate these specific behaviors into dedicated metrics. Nonetheless, to investigate performance across diverse driving contexts, we provide the following breakdown:

ContextGroupsNo feedbackReCoursive-SΔ [min, max]
Unsignalised right of way (stop sign, all-way stop, unprotected turn)5243.0347.69+4.65 [+4.36, +5.22]
Crosswalk / pedestrians8444.4348.35+3.92 [+3.22, +4.63]
Signalised intersection3945.2347.67+2.43 [+1.73, +2.86]
None of the above10449.4356.13+6.70 [+6.39, +7.29]
All22547.1252.47+5.34 [+4.92, +5.85]

Note: driving contexts can overlap.

Takeaway: the gain of latent feedback is not restricted to certain scenario types.

Revision: we have added these submetric and scenario breakdowns to Appendix D.3 (Tables 16 and 17).

Q4. Mechanism and novelty — targeted validation. Does the benefit persist when later planning gradients are detached at the latent feedback interface, while assessment supervision and forward feedback remain unchanged? This would help establish whether training the critic through its use in revision is consequential, and sharpen the distinction from a supervised feature-feedback refinement architecture.

We appreciate the insightful question! We ran the experiments with the stop-gradient setup, and compared it with explicit scores, no feedback, and ReCoursive-S.

Detaching later planning gradients removes the feedback benefit (47.05 vs. 47.12 for no feedback, compared with 52.47 for ReCoursive-S). All results below are cross-fit EPDMS on navhard (mean [min, max] over 3 fold seeds):

Variantdepth 1depth 2depth 6depth 6 - depth 1
Explicit scores47.26 [46.30, 48.25]47.27 [46.64, 47.95]46.79 [45.79, 47.45]-0.47 [-0.80, -0.09]
No feedback45.87 [45.04, 46.89]47.09 [46.67, 47.69]47.12 [46.55, 47.47]+1.25 [+0.57, +1.67]
Stop-gradient latent feedback46.03 [45.41, 46.77]46.87 [46.76, 47.00]47.05 [46.03, 47.97]+1.02 [+0.62, +1.23]
Latent feedback (ReCoursive-S)48.98 [48.72, 49.15]52.49 [52.43, 52.58]52.47 [52.39, 52.61]+3.49 [+3.33, +3.68]

Takeaway: Stop-gradient performs similarly to no feedback and substantially underperforms ReCoursive-S. The result aligns with our claim that the critic must learn from trajectory losses to guide revisions.

Revisions: We have added the stop-gradient ablation to Section 4.3 and Table 1 (they are currently single training seed, and we will train the stop-gradient with additional seeds for later revisions).

Part 3/3

Q5.1 Scaling and robustness — clarification of the supported claim. Are the main feedback gains and the six-versus-eight-step results reproducible across training seeds?

Yes: both the feedback gains and the 6-vs.-8-step ordering hold across all 3 training seeds with non-overlapping ranges. We trained two additional seeds for each configuration below and evaluated every checkpoint with 5-fold cross-fit over three random fold seeds. We report the mean [min, max] over 3 training seeds.

Variantdepth 1depth 2depth 6depth 8depth 6 - depth 1
Latent feedback (ReCoursive-S, 6 steps)48.58 [47.62, 49.13]49.73 [47.13, 52.49]52.16 [51.30, 52.72]—+3.58
ReCoursive-S, no feedback (6 steps)46.19 [45.24, 47.44]46.55 [44.61, 47.95]46.39 [44.54, 47.51]—+0.20
ReCoursive-S, explicit scores (6 steps)46.41 [45.73, 47.26]46.72 [46.23, 47.27]46.62 [46.02, 47.06]—+0.21
ReCoursive-S–8 (8 training steps)46.75 [46.02, 47.12]47.44 [46.75, 48.45]47.56 [46.83, 48.62]47.53 [46.87, 48.57]+0.81

Takeaways:

  1. Feedback gains hold: At depth 6, ReCoursive-S leads no feedback and explicit scores across all seeds, with a much larger revision gain from depth 1 to 6.
  2. Latent vs. explicit-score feedback: Latent feedback consistently outperforms explicit-score feedback across all seeds.
  3. 6 vs. 8 training steps: ReCoursive-S–8 consistently trails the 6-step model across all 3 seeds.
  4. Progression across depth: Under held-out selection, ReCoursive-S already leads no feedback at depth 1, and revisions from depth 1 to 6 add a further +3.58 EPDMS.

Revisions: Due to time constraints, we trained additional seeds for these four core configurations first. We will train the stop-gradient and MLP explicit-score with additional seeds for later revisions.

Q5.2 Have you evaluated inference beyond the training depth? Please clarify whether the intended claim is refinement within a trained unrolling horizon or generalizable test-time scaling; evidence for the latter would substantially strengthen the contribution.

Our intended claim is refinement within a trained unrolling horizon, not generalizable test-time scaling. When evaluating inference depths beyond the training horizon, we observe neither further gains nor performance drops; performance remains constant (mean EPDMS of cross-fit evaluations over 3 fold seeds):

Methoddepth 6depth 7depth 8depth 9depth 10depth 11depth 12depth 12 - depth 6
ReCoursive-S52.4852.4552.5152.5152.5152.5152.51+0.03

Note: coefficients are fitted on the depth-12 pool and shared across depths, so depth 6 differs slightly from 52.47.

To clarify, we do not claim test-time scaling in the paper. We acknowledge the limitation at line 525 that the revision gains concentrate in early depths (i.e., the revision benefits are largely realized in early depths). Beyond the trained horizon, extra steps neither improve EPDMS further nor outscore proposals from earlier steps, as extrapolating past the training depth lies outside our intended design.

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

ReCoursive extends a proposal-and-scoring driving planner with a recurrent feedback loop. A weight-shared proposer generates trajectories, and the critic's continuous pre-head representations condition subsequent revisions. Trajectory imitation and geometry-based assessment losses supervise every step, while later planning losses also train earlier critiques through the feedback path. Final selection considers the union of proposals across steps. The paper studies feedback versus recurrence alone, explicit scores versus latent feedback, assessment supervision, stale feedback, depth and latency. It reports NAVSIM v2 results and zero-shot evaluations on HUGSIM and KITScenes LongTail.

Critical strengths
  1. The paper asks a focused question and includes controls beyond a simple leaderboard comparison. Feedback and no-feedback models have the same proposal budget and recurrent depth. At six steps, the reported difference is 53.24 versus 50.41 EPDMS; the four-step experiment has the same direction. Stale and zero-feedback interventions retain current scores for selection, helping distinguish using feedback for revision from simply rescoring candidates. These are meaningful controls for the proposed mechanism, subject to the evaluation concern below.

  2. The work includes useful transfer evidence and unusually explicit implementation details. HUGSIM uses uniform coefficients rather than the NAVSIM coefficient search, and the four-step model achieves 52.5 route completion and 43.6 HD-Score across 345 scenarios. The appendix specifies proxy target construction, gradient detachment, borrowed-trajectory supervision, adapters and timing. The authors appropriately identify the latent-editing examples as selected illustrations, not evidence of reliable directional control.

Critical weaknesses
  1. The primary benchmark is also used to fit the final decision rule. Appendix C.5, p.17, states that each checkpoint's selection coefficients are optimized on all 5,912 navhard evaluation scenes in 225 reactive groups, using 16,807 coarse vectors plus 3,124 refinements. The resulting EPDMS is then reported on those same scenes. Although network weights remain frozen, the selector is part of the planner, so this is not an independent estimate of generalization. Refitting some released baselines is a helpful audit of tuning sensitivity, but does not remove this problem. Moreover, TOAD is evaluated with its released coefficients or one transferred vector rather than an equivalent search of its complete CEM-and-selection pipeline. Consequently, neither the approximately 1-point lead over TOAD nor the magnitude of the per-model ablation gaps is yet established under held-out selection. This concern does not negate the separate HUGSIM transfer results.

  2. Several central explanations are supported by single-run comparisons without isolating all of the claimed learning mechanism. Appendix B.1 specifies seed 2 for NAVSIM-only training, and the main feedback/depth tables give no training-seed variation or uncertainty. The explicit-score model starts 1.75 points higher at depth one and loses 0.49 points across depth, while the eight-step training run trails the six-step run at every depth. Replication is particularly important for interpreting these as interface or depth effects. In addition, removing assessment losses changes the shared representation and borrowed-trajectory training, not only the semantic grounding of the feedback; the authors correctly acknowledge this in Appendix D.2. A stop-gradient control on the critique-to-proposer feedback, with the remaining forward computation and losses retained, would directly test the distinctive claim that training the critic through its downstream use is important. The existing results support a promising design, but not yet all of the broader mechanistic interpretation.

Questions for authors
  1. Can you fit all selector coefficients on a separate validation set and report the main comparisons and ablations on untouched reactive groups? If only navhard is available, can you provide group-disjoint cross-fitting, preserving complete first/second-stage groups and separating source logs where possible? In particular, do the feedback advantage and lead over TOAD persist?
  2. Across multiple training seeds, how stable are feedback versus no feedback, latent versus score feedback, and six versus eight training steps? Please report paired uncertainty over evaluation groups in addition to training variation.
  3. Does stopping gradients through the critique feedback, while retaining recurrence and all assessment losses, change the benefit? For the explicit-score comparison, would a nonlinear projection or logit-based interface alter the conclusion that predicted scores fail to support revision?
3 comments

Part 1/2

We thank reviewer bzZK for the constructive review and the effort invested. We are glad to address the concerns raised.

Q1. Can you fit all selector coefficients on a separate validation set and report the main comparisons and ablations on untouched reactive groups? If only navhard is available, can you provide group-disjoint cross-fitting, preserving complete first/second-stage groups and separating source logs where possible? In particular, do the feedback advantage and lead over TOAD persist?

Yes: both the feedback advantage (+5.4 EPDMS) and our lead over TOAD (+2.3 EPDMS) persist. Our group-disjoint 5-fold cross-fitting preserves complete first/second-stage groups. Please see the common response above for the full comparison table, and Sections 4.2–4.4 of the revised paper, where all ablations now also report cross-fit EPDMS.

Q2. Across multiple training seeds, how stable are feedback versus no feedback, latent versus score feedback, and six versus eight training steps? Please report paired uncertainty over evaluation groups in addition to training variation.

All three comparisons hold consistently across 3 training seeds, with non-overlapping ranges at depth 6.

Across the 225 evaluation groups (paired bootstrap 95% CI on the original seed), the depth-6 differences are +5.34 [+2.72, +7.94] over no feedback, +5.68 [+2.97, +8.44] over explicit scores, and +3.90 [+1.35, +6.49] over ReCoursive-S–8 (depth 8).

We trained two additional seeds for each configuration below and evaluated every checkpoint with 5-fold cross-fitting (averaged over the 3 random 5-fold partitions). We report the mean [min, max] over 3 training seeds.

Variantdepth 1depth 2depth 6depth 8depth 6 - depth 1
Latent feedback (ReCoursive-S, 6 steps)48.58 [47.62, 49.13]49.73 [47.13, 52.49]52.16 [51.30, 52.72]—+3.58
ReCoursive-S, no feedback (6 steps)46.19 [45.24, 47.44]46.55 [44.61, 47.95]46.39 [44.54, 47.51]—+0.20
ReCoursive-S, explicit scores (6 steps)46.41 [45.73, 47.26]46.72 [46.23, 47.27]46.62 [46.02, 47.06]—+0.21
ReCoursive-S–8 (8 training steps)46.75 [46.02, 47.12]47.44 [46.75, 48.45]47.56 [46.83, 48.62]47.53 [46.87, 48.57]+0.81

Takeaways:

  1. Feedback gains hold: At depth 6, ReCoursive-S leads no feedback and explicit scores across all seeds, with a much larger revision gain from depth 1 to 6.
  2. Latent vs. explicit-score feedback: Latent feedback consistently outperforms explicit-score feedback across all seeds.
  3. 6 vs. 8 training steps: ReCoursive-S–8 consistently trails the 6-step model across all 3 seeds.
  4. Progression across depth: Under held-out selection, ReCoursive-S already leads no feedback at depth 1, and revisions from depth 1 to 6 add a further +3.58 EPDMS.

Revisions: Due to time constraints, for now we trained only the above configs with additional training seeds. We will train the stop-gradient and MLP explicit-score with additional seeds for later revisions.

Part 2/2

Q3. Does stopping gradients through the critique feedback, while retaining recurrence and all assessment losses, change the benefit? For the explicit-score comparison, would a nonlinear projection or logit-based interface alter the conclusion that predicted scores fail to support revision?

Stopping gradients removes the feedback benefit (47.05 vs. 47.12 for no feedback), and neither logit (46.99) nor MLP (50.11) score interfaces match latent critiques (52.47). All the results are cross-fit EPDMS on navhard with 3 fold seeds.

Variantdepth 1depth 2depth 6depth 6 - depth 1
Explicit scores (paper, affine projection)47.26 [46.30, 48.25]47.27 [46.64, 47.95]46.79 [45.79, 47.45]-0.47 [-0.80, -0.09]
Explicit scores, logit input45.69 [45.25, 46.25]45.95 [45.49, 46.23]46.99 [46.68, 47.53]+1.31 [+1.13, +1.52]
Explicit scores, MLP projection49.43 [49.24, 49.80]50.79 [50.65, 50.97]50.11 [49.99, 50.20]+0.68 [+0.33, +0.97]
Stop-gradient latent feedback46.03 [45.41, 46.77]46.87 [46.76, 47.00]47.05 [46.03, 47.97]+1.02 [+0.62, +1.23]
No feedback45.87 [45.04, 46.89]47.09 [46.67, 47.69]47.12 [46.55, 47.47]+1.25 [+0.57, +1.67]
Latent feedback (ReCoursive-S)48.98 [48.72, 49.15]52.49 [52.43, 52.58]52.47 [52.39, 52.61]+3.49 [+3.33, +3.68]

Takeaways:

  1. Stop-gradient performs similarly to no feedback and substantially underperforms ReCoursive-S. The result aligns with our claim that the critic must learn from trajectory losses to guide revisions.
  2. Explicit-score variants yield at most the recurrence gain of no feedback.
  3. ReCoursive-S achieves both higher final EPDMS and substantially larger revision gains than all explicit-score variants.

Revisions: We have added the explicit-score variants (details in Appendix D.1) and the stop-gradient ablation to Section 4.3 and Table 1. Due to time constraints, these results use a single training seed. We will train the stop-gradient and MLP explicit-score with additional seeds for later revisions.

Dear Reviewer bzZK, thank you again for your feedback. With the discussion period closing in about 24 hours, we wanted to check whether you have had a chance to review our responses above and the revised manuscript. Please let us know if any questions remain while we still have time to reply.

Comments on the paper

4 comments

Response to fitting selection coefficients on the testset navhard (bzZK Q1, QBm8 Q1)

We appreciate that reviewers raised this concern. We considered this evaluation trade-off while preparing the paper.

No prior work evaluated on navhard fits coefficients disjointly from navhard or releases search code.

Previous works either do not state which split is used to fit the coefficients [1–4], or reuse the scorer and coefficients of their base model in the released code [5, 6].

DrivoR [7] reports tuning on the warmup split, which remains a subset of navhard. More generally, the released coefficients of DrivoR, DrivoR+SimScale, and SimScale rank within the top 0.2% of this search on navhard (see Table 13).

To our knowledge, we are the first to investigate the baseline coefficients and disclose the full search procedure used in the method (see Appendix C.5).

Rationale for full-testset fitting.

We weighed two options:

  • Full-set fitting enables direct comparison with published numbers, including methods without released checkpoints.
  • Cross-fitting alone limits the comparison to methods we can re-run, and introduces a protocol that no published paper uses.

We prioritized ensuring fair and direct comparability with published baselines. We therefore chose full-split fitting, described it explicitly, and applied our search to the released baselines to keep the comparison fair.

Our cross-fit results.

We evaluate models with group-disjoint 5-fold cross-fitting, repeated over 3 random 5-fold partitions (fold seeds). For each of the five folds, we search coefficients on 4 folds (180 reactive groups) and evaluate on the held-out fold (45 reactive groups). We average over the five held-out folds and report the mean [min, max] across 3 fold seeds.

For TOAD, searching ~20,000 vectors directly inside its online CEM loop is computationally prohibitive, as every candidate vector requires new CEM rollouts and simulator evaluations across scenes. Because TOAD's released code reuses DrivoR's coefficients directly, we follow the same practice: in each fold, we tune DrivoR's coefficients on the 180 fitting groups and run TOAD's CEM with them on the 45 held-out groups.

MethodTraining regimeEPDMS (Table 2)cross-fit EPDMS
ZTRSNAVSIM48.147.0 [46.2, 47.7]
DrivoRNAVSIM48.344.0 [42.6, 45.1]
Gigapixel-DrivoRSelf-play + NAVSIM adaptation50.1not open-sourced yet
ReCoursive-SNAVSIM53.252.5 [52.4, 52.6]
ReCoursive-S, no feedbackNAVSIM50.447.1 [46.6, 47.5]
SimScaleNAVSIM + SimScale53.2†50.2 [49.1, 50.8]
DrivoRNAVSIM + SimScale54.552.8 [52.2, 53.4]
TOADNAVSIM + SimScale56.353.8 [52.9, 54.4]
ReCoursive-LNAVSIM + SimScale57.256.1 [55.9, 56.2]

† Leaderboard score that we could not reproduce. The released checkpoint scores 51.3 on the full set (Table 13).

Takeaways:

  1. ReCoursive-S vs. ReCoursive-S, no feedback: The feedback advantage holds, with ReCoursive-S outperforming the no-feedback model by 5.4 EPDMS.
  2. ReCoursive-S remains the best among all compared NAVSIM-trained models, and ReCoursive-L achieves the highest score among compared methods.
  3. ReCoursive-S and L show the best robustness to fold partitioning (varying by at most 0.3 EPDMS among the 3 fold seeds).

Revisions

We have extended Appendix C.5 with our cross-fit protocol and results (Table 12), and Appendix C.6 with the coefficient selection of other works. Sections 4.2–4.4 of the revised paper now report cross-fit EPDMS; the benchmark comparison in Section 4.5 keeps the full-split protocol of prior work, with cross-fit counterparts in Table 12. Regarding whether to adopt the cross-fit results as the primary metric for the benchmark comparison as well, we welcome the reviewers' feedback and opinions. Upon acceptance, we will also release our code, model weights, and coefficient-search scripts.

[1] Z. Li et al. Generalized Trajectory Scoring for End-to-End Multimodal Planning. arXiv:2506.06664, 2025.

[2] Z. Li et al. ZTRS: Zero-Human Demonstration End-to-End Autonomous Driving with Trajectory Scorer. ECCV 2026.

[3] L. Rowe et al. Scaling Self-Play for End-to-End Driving. CoRL 2026.

[4] Z. Li et al. Hydra-MDP: End-to-End Multimodal Planning with Multi-Target Hydra-Distillation. arXiv:2406.06978, 2024.

[5] H. Tian et al. SimScale: Learning to Drive via Real-World Simulation at Scale. CVPR 2026.

[6] Y. Xu et al. Test-Time Trajectory Optimization for Autonomous Driving. arXiv:2606.07170, 2026.

[7] E. Kirby et al. Driving on Registers. CVPR 2026.

Overview

We thank the reviewers (bzZK, QBm8, TfvF) for their constructive feedback. All three reviewers valued our research question on what should pass from assessment to revision, and bzZK and QBm8 also highlighted our controlled ablations and transparent reporting.

Across the reviews, the main concerns center on clear, testable questions: held-out selector fitting, stability across training seeds, and ablations isolating our learning mechanism. We ran new experiments on each, and the results support our main claims:

  1. Held-out selection (bzZK, QBm8). Under 5-fold cross-fitting, ReCoursive-S leads its no-feedback control by 5.4 EPDMS (52.5 vs. 47.1), and ReCoursive-L leads TOAD by 2.3 EPDMS (56.1 vs. 53.8).
  2. Mechanism (bzZK, QBm8, TfvF). In single-seed runs, stopping gradients at the critique interface drops the model to the no-feedback level (47.05 vs. 47.12), well below ReCoursive-S (52.47). Logit and MLP score interfaces also stay lower (46.99 and 50.11).
  3. Training seeds (bzZK, QBm8). Across three training seeds, ReCoursive-S averages 52.16, compared with 46.39 (no feedback), 46.62 (explicit scores), and 47.53 (eight training steps).
  4. Further analyses (QBm8, TfvF). We have added sub-metric and scenario breakdowns to the paper, and we report inference up to depth 12 and editing statistics across 66 and 85 scenes.

Dear Reviewers,

Less than two days remain before the discussion period ends on November 18 (AoE). If you have not yet replied to the author rebuttal, please do so as soon as possible, indicating which of your concerns are resolved and which remain.

Best regards, Area Chair

AuthorsOpenReview ↗

Summary of our paper and rebuttal

Dear AC, SAC, PCs, and Reviewers,

As the discussion period closes, we would like to share a self-contained summary of our paper and rebuttal. Despite low initial scores, we chose not to withdraw this submission. We stand behind both our research question, which all three reviewers named as a strength, and how our experiments answer it. The reviews centered on clear, testable scientific questions. We tested each one, and the results support our claims.

What the paper does. Many end-to-end driving planners propose candidate trajectories, rate each with a learned scorer, and drive the best-rated one. The scorer's judgment is then discarded, so if every candidate is flawed, the planner can only pick the least flawed one. ReCoursive lets the planner try again. Its scorer acts as a critic: for each candidate, it hands back the internal features from which it computes its scores. We call these latent critiques. The proposer reads them and generates revised candidates, for up to six rounds with shared weights.

What is new. To our knowledge, what feedback makes revision useful has so far been studied only for language, where a model writes a critique as text and then revises. In driving, language-based reasoning means training on reasoning traces and generating text at run time. ReCoursive generates no text and trains only on demonstrated trajectories and score labels, the supervision that scorer-based planners already use. Four rounds add 8.4 ms over one. Earlier driving planners score candidates without revising them (Hydra-MDP, DrivoR), revise without an assessment (iPad, DiffRefiner), or pass scalar scores to a hand-designed search (TOAD). ReCoursive learns the revision and trains the critic for how its feedback is used. The same question arises wherever a learned scorer ranks generated candidates.

What a reader learns. Scores are EPDMS, the NAVSIMv2 driving score (0 to 100), on the benchmark's hard split.

  • Feedback helps beyond extra computation. The same model without feedback, with equal rounds and candidates, scores 47.1. With latent critiques it scores 52.5.
  • What is passed matters. Passing predicted scores instead of features, in three variants, ends at 46.8 to 50.1, and extra rounds add about as much as without feedback (-0.5 to +1.3, against +1.3). With latent critiques, extra rounds add +3.5.
  • How the critic is trained matters. If it is trained only to predict scores, and not through the revisions that use its critiques, the score returns to 47.1.
  • Critiques must be current. Reusing first-round critiques gives 50.9, and zeroing them gives 48.7.

Results. On NAVSIMv2, ReCoursive leads the compared learned methods: 53.2 without extra data (its base model DrivoR: 48.3, next best: 50.1) and 57.2 with SimScale training data (TOAD: 56.3). Without fine-tuning on the target benchmarks, it also achieves the highest route completion and HD-Score among compared methods in closed-loop driving on HUGSIM, with an HD-Score 13% above the previous best, and the highest score among methods without a vision-language model on KITScenes LongTail.

How we addressed the reviews. The reviewers shared three requests, and the conclusions held under each.

  • Held-out selection. Following the convention on NAVSIMv2, we had tuned the five weights that combine the critic's scores on the evaluation split. We now fit them on four fifths of the scenes and evaluate on the rest. All numbers in the list above use this protocol. Under it, ReCoursive scores 52.5 and 56.1, against 47.0 and 53.8 for the best learned baselines we could re-run.
  • Training seeds. Over three seeds, ReCoursive averages 52.2 against 46.4 without feedback, with non-overlapping ranges.
  • Mechanism ablations. The critic-training experiment and two of the three score variants above were run at the reviewers' request.

The reviewers' other points are answered in our individual responses and the revised paper.

Reproducibility. Reviewers bzZK and QBm8 listed the paper's explicit and transparent reporting among its strengths. The release will follow the same standard: code, model weights, and our coefficient-search scripts, upon acceptance. To our knowledge, none of the compared methods releases such a search, so others can then adopt held-out selection as well.

We thank the reviewers for their feedback, and the chairs for handling our submission.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.