acceptodds
Under review as a conference paper at ICLR 2027

A-Flow: Perturbation Can Be Correction for Frozen Flow Planners in End-to-End Driving

Abstract

Recent end-to-end driving methods use generative models for trajectory planning. They apply reinforcement learning (RL) fine-tuning to improve driving safety and comfort. However, such fine-tuning can be costly and sensitive to training data. Parameter updates may also disrupt pretrained driving behavior. Inspired by non-reversible Langevin dynamics, we design a lightweight skew-symmetric matrix, denoted by , to parameterize a structured perturbation that is learned as a trajectory correction. Building on this design, we propose A-Flow, a driving framework with three components. First, Temporal BEV Mapping (TBM) combines camera history to represent road geometry and traffic dynamics for trajectory generation. Second, we use GRPO to search for high-reward trajectories around the uncorrected trajectory. We optimize only while keeping the pretrained planner frozen. The learned matrices form a correction library for reuse. Third, Field-Guided Correction Transfer (FCT) retrieves relevant matrices to generate corrected candidates in unseen scenes. It uses TBM features to evaluate these candidates and select the final trajectory. A-Flow achieves 90.7 PDMS in NAVSIM v1 and an HD-Score of 37.49 in HUGSIM. HUGSIM evaluation uses zero-shot transfer without target-scene reward optimization. Experiments on a ReCogDrive-based planner further demonstrate the adaptability of our correction method.

Primary area:
applications to robotics, autonomy, planning
Keywords:
end-to-end autonomous driving, flow matching, reinforcement learning, trajectory correction, zero-shot transfer

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_JATsOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

A-Flow improves a frozen flow-based driving planner using learned skew-symmetric correction matrices. These matrices modify the final sampling step and are stored in a library for reuse across scenes. A retrieval and selection pipeline chooses suitable corrections at inference time.

Critical strengths

The experiments are extensive, covering two driving benchmarks, component ablations, and an additional flow-based planner. The reported improvements suggest that learned corrections can be useful across different scenes and planner backbones.

Critical weaknesses
  1. The paper is difficult to follow. The core role of the A matrix should be explained more directly before introducing the theoretical motivation and the full selection pipeline. Some implementation and training details could be moved to the experimental section or appendix.
  2. The novelty of the correction mechanism needs stronger justification. At a high level, A is a learned transformation that adjusts an initial trajectory proposal. It remains unclear what advantage its skew-symmetric structure offers over a small MLP or another learned residual correction module.
  3. The practical efficiency benefit is not fully established. Although A is lightweight, learning the correction library still requires reward-based optimization. So how's the time consuming with the GRPO uses on the total trajectory decoder.
Questions for authors
  1. What distinguishes the proposed correction from an MLP-based trajectory refinement module? Could you compare them under the same frozen planner, reward budget, and candidate-selection pipeline?
  2. How many trainable parameters does each A matrix require, and how large is the complete correction library relative to the trajectory decoder?
  3. What is the total training cost, including matrix optimization, library construction, and evaluator training? How does it compare with applying GRPO directly to the trajectory decoder at comparable performance?
6 comments
AuthorsOpenReview ↗

How the correction matrix works

Thank you for your suggestions. Matrix AA corrects the final sampling step, with skew symmetry constraining the correction direction. The correction bank outperforms a similarly sized shared MLP, and A-Flow outperforms direct Flow-GRPO at comparable training cost, including evaluator training.

How the correction matrix works

W1. The paper is difficult to follow. The core role of the A matrix should be explained more directly before introducing the theoretical motivation and the full selection pipeline. Some implementation and training details could be moved to the experimental section or appendix.

The correction matrix AA is a 24×2424 \times 24 linear operator applied at the frozen planner's final sampling step, allowing multiple candidates to share one flow rollout. It maps the endpoint proxy q⋆q_\star to the correction term ρAq⋆\rho A q_\star, applied only at t⋆=0.8t_\star = 0.8. Candidates share the first four steps and the frozen network output at the final step.

Training stores the correction directions learned for each source scene as a matrix. At test time, retrieval supplies relevant matrices, each matrix proposes a correction to the current trajectory, and FCT assesses whether it suits the target scene. We now explain this role of AA at the start of Section 4 and present implementation settings in Section 5, with detailed conditioning, optimization, and bank-construction procedures in Appendices A, C.2, and D.1.

AuthorsOpenReview ↗

Skew symmetry and comparison with a residual MLP

Skew symmetry and comparison with a residual MLP

W2. The novelty of the correction mechanism needs stronger justification. At a high level, A is a learned transformation that adjusts an initial trajectory proposal. It remains unclear what advantage its skew-symmetric structure offers over a small MLP or another learned residual correction module.

Q1. What distinguishes the proposed correction from an MLP-based trajectory refinement module? Could you compare them under the same frozen planner, reward budget, and candidate-selection pipeline?

Skew symmetry improves matrix corrections with fewer degrees of freedom, and the correction bank outperforms a similarly sized shared MLP. Table 4 compares unconstrained and skew-symmetric matrices under full A-Flow (including scale adjustment and the second branch); the MLP comparison below evaluates 32-candidate single-branch selection (A-Flow32) alongside full A-Flow.

Matrix parameterizationIndependent degrees of freedomOracle@32Final PDMS
Learned unconstrained matrix MM57693.1888.81
Learned skew-symmetric matrix AA27695.1390.69

The higher scores support the directional constraint, which has a direct geometric interpretation: in normalized trajectory coordinates, before clipping, A=B−B⊤A = B - B^\top guarantees q⋆⊤Aq⋆=0q_\star^\top A q_\star = 0. A perturbation along Aq⋆Aq_\star therefore produces zero first-order change in the Euclidean norm of q⋆q_\star (Appendix B.3).

The correction bank achieves higher oracle and final scores than a shared residual MLP. We reproduce CLEAR on NAVSIM, adapting its trajectory-refinement module to the same frozen planner used by A-Flow. Our implementation uses a 7.94M-parameter shared residual MLP, trained with PPO and vehicle-state feedback until validation performance on the monitoring split plateaued. Its size is comparable to the deployed bank's 7.35M dense elements. The results below are from our NAVSIM reproduction of CLEAR. With 32 candidates and the same single-branch Value+Safety FCT pipeline, A-Flow32 scores 3.50 PDMS points higher across all 12,146 navtest scenes.

Correction systemCorrector sizeNumber of corrected candidatesCorrected-candidate Oracle@32Final PDMS
CLEAR (our NAVSIM reproduction, deterministic output)MLP: 7.94M parameters1—85.20
CLEAR (our NAVSIM reproduction) + FCTMLP: 7.94M parameters3291.1686.81
A-Flow32Deployed correction bank: 7.35M dense elements3295.1390.31
Full A-FlowDeployed correction bank: 7.35M dense elements2×32—90.70

The higher oracle score shows that the bank's advantage begins before selection. Similar ego states may require opposite avoidance directions when road geometry or obstacles differ. A lightweight shared MLP may compromise between these modes. The discrete bank instead preserves correction directions from individual source scenes for retrieval and selection in each target scene, helping retain candidate diversity and quality.

AuthorsOpenReview ↗

Matrix and correction-bank size

Matrix and correction-bank size

Q2. How many trainable parameters does each A matrix require, and how large is the complete correction library relative to the trajectory decoder?

The deployed bank contains roughly one quarter as many dense elements as the full Stage2 decoder has parameters; optimization updates only 576 matrix parameters per scene. Optimized parameters refer to BB; independent degrees of freedom refer to A=B−B⊤A = B - B^\top. FP32 storage counts only weights or matrices.

ItemParameters / dense elementsIndependent degrees of freedomSize relative to decoderFP32 storage
Full Stage2 trajectory decoder (frozen)30,876,675—100%117.79 MiB
Single 24×2424 \times 24 matrix5762760.0019%2.25 KiB
Full unfiltered set, 85,109 source matrices49.02M23.49M158.77%Approx. 187.0 MiB
Deployed correction bank, 12,767 matrices7.35M3.52M23.82%28.05 MiB
32 matrices retrieved by one branch18,4328,8320.06%72 KiB
64 matrices retrieved by two branches36,86417,6640.12%144 KiB

The bank stores reusable scene corrections. Training maintains gradients and optimizer states only for active matrices, and inference applies only the retrieved subset. Matrix computation at each update or generation step therefore depends on the matrices in use, rather than the full bank size.

Reviewer_TYv3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

A-Flow improves a frozen flow-based driving planner using learned skew-symmetric matrices applied at the final sampling step. The matrices form a correction library retrieved using scene and trajectory descriptors. Learned value and safety evaluators select among corrected candidates, with additional correction-strength adjustment and retrieval branches. Experiments demonstrate improvements on NAVSIM, zero-shot transfer to HUGSIM, and adaptation to a ReCogDrive-based flow planner.

Critical strengths
  1. Innovative formulation with substantial theoretical analysis. Reusing structured terminal corrections offers an appealing alternative to full-planner fine-tuning. The theoretical analysis motivates the skew-symmetric parameterization and characterizes the geometry and bounded deviation of the implemented correction.

  2. Strong empirical improvements. NAVSIM PDMS increases from 85.4 to 90.7, and HUGSIM HD-Score from 24.26 to 37.49. The improved oracle and selected scores over learned unconstrained matrices further support the proposed parameterization.

Critical weaknesses
  1. The source of the gains remains insufficiently isolated. Table 4 supports the matrix design, but lacks comparisons against ordinary flow or diffusion candidates paired with an equivalently trained scorer. Aggregate performance for a single retrieved correction without post-selection is also missing. These results are important for distinguishing reliable correction transfer from effective candidate generation and selection.

  2. The inference trade-off is less convincing. Full-system latency rises from 66.61 to 241.33 ms, while the value-and-safety configuration already reaches 90.3 PDMS at 125.02 ms. A runtime breakdown and a clearer performance–latency comparison are needed to establish the practical advantage of the retrieval and selection pipeline.

Questions for authors
  1. Could comparable gains be achieved using directly sampled flow candidates, stochastic diffusion rollouts as used in ReCogDrive’s RL stage [1], or DiffusionDrive-style anchors with truncated diffusion [2], paired with an equivalently trained scorer? Please compare oracle and selected performance under matched candidate counts.

  2. How much inference time is spent on retrieval, candidate feature extraction, and the successive selection stages? Can fewer retrieved candidates or fewer selection passes retain most of the gains at substantially lower latency?

[1] Li et al. ReCogDrive: A Reinforced Cognitive Framework for End-to-End Autonomous Driving. ICLR, 2026.

[2] Liao et al. DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving. CVPR, 2025.

9 comments
AuthorsOpenReview ↗

Comparison with flow and diffusion candidates

Thank you for your suggestions. Matrix corrections improve the candidate pool, and FCT selects useful corrections for each target scene. With candidates parallelized within a branch, inference cost depends mainly on repeated selection and feature re-evaluation, and on whether the second retrieval branch is enabled.

Comparison with flow and diffusion candidates

W1 (candidate comparisons). The source of the gains remains insufficiently isolated. Table 4 supports the matrix design, but lacks comparisons against ordinary flow or diffusion candidates paired with an equivalently trained scorer.

Q1. Could comparable gains be achieved using directly sampled flow candidates, stochastic diffusion rollouts as used in ReCogDrive’s RL stage [1], or DiffusionDrive-style anchors with truncated diffusion [2], paired with an equivalently trained scorer? Please compare oracle and selected performance under matched candidate counts.

References cited in Q1. [1] Li et al. ReCogDrive: A Reinforced Cognitive Framework for End-to-End Autonomous Driving. ICLR, 2026.

[2] Liao et al. DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving. CVPR, 2025.

With imitation-learning backbones, learned corrections achieve higher final PDMS than flow or diffusion sampling followed by FCT, at low candidate generation cost. Each method uses 32 candidates. FCT is retrained on each method's candidates with the same architecture and training budget, including ReCogDrive-IL32 and ReCogDrive-RL32. Across all 12,146 navtest scenes, A-Flow32 outperforms Independent Flow32 and DiffusionDrive32 by 2.49 and 0.58 PDMS points, respectively.

MethodCandidate generation (GFLOPs)Oracle@32Final PDMS
Independent Flow321,181.1791.0387.82
ReCogDrive-IL322,308.9189.9188.50
DiffusionDrive3239.1296.8489.73
A-Flow3237.0295.1390.31
ReCogDrive-RL322,308.9192.4491.20

DiffusionDrive has a higher oracle score, but its diffusion candidates generated from static anchors average only 59.65, versus 81.86 for A-Flow32. This suggests a harder selection problem even after FCT retraining. ReCogDrive-RL32, which uses the official RL-fine-tuned planner checkpoint, scores 0.89 PDMS points above A-Flow32 (0.50 above full A-Flow) while requiring 62.4 times the candidate-generation GFLOPs of A-Flow32.

The comparison using the same Flow backbone shows both benefits most clearly: correction directions learned from rewards raise the candidate ceiling, while the shared prefix over the first four steps and the shared frozen network output at the final step reduce repeated generation. FCT evaluates the resulting trajectories for value and safety in the target scene.

AuthorsOpenReview ↗

The role of selection

The role of selection

W1 (single correction). Aggregate performance for a single retrieved correction without post-selection is also missing. These results are important for distinguishing reliable correction transfer from effective candidate generation and selection.

Selection in the target scene helps realize gains from retrieved corrections. Directly using the top-ranked correction lowers mean PDMS. For this comparison, we fix ρ=0.7\rho = 0.7 and disable FCT and the Base fallback.

Evaluation setBaseDirect Rank-1 outputDifference
navtest, 12,146 scenes85.3882.52−2.86

Retrieval measures source–target similarity, which need not reflect the quality of the resulting correction. FCT uses predicted value and safety in the target scene to choose among corrected candidates and Base.

AuthorsOpenReview ↗

Inference cost and performance trade-off

Inference cost and performance trade-off

W2. The inference trade-off is less convincing. Full-system latency rises from 66.61 to 241.33 ms, while the value-and-safety configuration already reaches 90.3 PDMS at 125.02 ms. A runtime breakdown and a clearer performance–latency comparison are needed to establish the practical advantage of the retrieval and selection pipeline.

Q2. How much inference time is spent on retrieval, candidate feature extraction, and the successive selection stages? Can fewer retrieved candidates or fewer selection passes retain most of the gains at substantially lower latency?

The single-branch Value+Safety configuration retains approximately 93% of full A-Flow's PDMS gain over Base while roughly halving three-stage processing time. The following results use the original evaluation configuration (B=1B=1) on the full navtest set.

ConfigurationPDMSThree-stage processing time
Base85.3866.61 ms
Value+Safety, A-Flow3290.31125.02 ms
Full A-Flow90.70241.33 ms

Full A-Flow additionally applies scale adjustment and a second retrieval branch, improving PDMS from 90.31 to 90.70 at additional computational cost. The single-branch configuration therefore provides an option when inference efficiency is a priority.

A separate concurrent benchmark quantifies the cost of adding each component. We use one A100 80GB with eight processes and B=8B=8 per process, keeping 32 candidates per retrieval branch. The configurations below add components cumulatively. Measurements include input processing and report amortized time per scene and aggregate throughput, rather than single-request latency.

Pipeline configuration (cumulative)Amortized concurrent time (ms/scene)Throughput (scenes/s)
Base29.6833.69
+ Value evaluation34.9628.61
+ Safety evaluation, A-Flow3236.9727.05
+ Scale adjustment42.9423.29
+ Second retrieval branch, full A-Flow53.6418.64

Omitting scale adjustment and the second retrieval branch reduces amortized inference time from 53.64 to 36.97 ms/scene, a 31.1% reduction. The PDMS results in the first table were measured in the original evaluation configuration, separately from this concurrent benchmark.

Candidate feature extraction and value/safety evaluation account for most of the processing cost. In the original configuration, retrieval and matrix transformations together take less than 1 ms. Within each branch, the 32 candidates are evaluated together as a batch. Scale adjustment and the second retrieval branch introduce additional feature extraction and candidate evaluation, explaining the higher cost of the full configuration.

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

This paper proposes A-Flow, a framework that improves a frozen flow-based planner through learned trajectory corrections. It learns scene-specific correction matrices and stores them in a library for retrieval and reuse in unseen scenes. The method achieves 90.7 PDMS on NAVSIM and an HD-Score of 37.49 in zero-shot evaluation on HUGSIM.

Critical strengths

1.The learned correction matrices can be retrieved and reused across scenes, providing a practical way to improve a frozen planner without updating its parameters.

2.The method improves NAVSIM PDMS from 85.4 to 90.7 and achieves an HD-Score of 37.49 in zero-shot evaluation on HUGSIM. These results provide evidence for the effectiveness of the correction and selection pipeline.

Critical weaknesses

1.It is not clear how much scene-based retrieval helps. Table 4 compares learned matrices with randomly generated matrices. However, it does not compare scene-based retrieval with randomly choosing the same number of matrices from the learned bank. The value/safety evaluation and refinement steps are quite complex, so it is hard to tell how much gain comes from retrieval itself.

2.The contribution of kinematic post-processing is unclear. Appendix D.3 includes trajectory scaling, lateral shifts, smoothing, blending with Base, and a braking fallback. These steps can directly change the final trajectory, but their contribution is not separately evaluated. It is therefore difficult to tell how much of the performance gain comes from the learned corrections and how much comes from post-processing.

Questions for authors

1.What is the result if you randomly choose 32 matrices from the same learned bank and keep the selection and refinement steps unchanged?

2.How much gain comes from the kinematic post-processing? Could you report the result without it?

5 comments
AuthorsOpenReview ↗

Retrieval versus random sampling from the same bank

Thank you for your suggestions. Retrieval improves candidate quality, and the main planning gains remain without post-processing. Kinematic post-processing further improves closed-loop execution.

Retrieval versus random sampling from the same bank

W1. It is not clear how much scene-based retrieval helps. Table 4 compares learned matrices with randomly generated matrices. However, it does not compare scene-based retrieval with randomly choosing the same number of matrices from the learned bank. The value/safety evaluation and refinement steps are quite complex, so it is hard to tell how much gain comes from retrieval itself.

Q1. What is the result if you randomly choose 32 matrices from the same learned bank and keep the selection and refinement steps unchanged?

Retrieval improves candidates before FCT selection and outperforms random sampling from the same bank. Both methods use 32 matrices from the same bank of 12,767 learned matrices. For Random A32, we retrain FCT on its own candidates with the same architecture and training budget. Retrieval improves PDMS by 1.21 across all 12,146 navtest scenes.

Candidate sourceCandidatesMean candidate PDMSOracle@32Final PDMS
Random A32, uniform sampling from the same bank3278.9893.4689.10
A-Flow32, RRF retrieval3281.8695.1390.31

Both use the same single-branch Value+Safety pipeline, without scale adjustment or the second retrieval branch. Adding scale adjustment and the second retrieval branch gives full A-Flow at 90.70 PDMS, as reported in Table 5.

The higher candidate mean and oracle score show that retrieval helps before final selection. Matching scene and trajectory features gives FCT corrections better suited to the current driving conditions.

AuthorsOpenReview ↗

Contribution of kinematic post-processing

Contribution of kinematic post-processing

W2. The contribution of kinematic post-processing is unclear. Appendix D.3 includes trajectory scaling, lateral shifts, smoothing, blending with Base, and a braking fallback. These steps can directly change the final trajectory, but their contribution is not separately evaluated. It is therefore difficult to tell how much of the performance gain comes from the learned corrections and how much comes from post-processing.

Q2. How much gain comes from the kinematic post-processing? Could you report the result without it?

A-Flow's main gains persist without kinematic post-processing. We remove post-processing while keeping the bank, retrieval, and FCT fixed. Across all 436 HUGSIM scenes, A-Flow outperforms Base by 10.02 HD-Score points with post-processing disabled for both.

Benchmark and metricMethodWithout post-processingWith post-processingGain from post-processing
NAVSIM, PDMSA-Flow90.691790.6950+0.0033
HUGSIM, HD-ScoreBase24.2625.49+1.23
HUGSIM, HD-ScoreA-Flow34.2837.49+3.21
HUGSIM, HD-ScoreA-Flow gain over Base+10.02+12.00—

Reliable trajectory execution matters more in HUGSIM's closed loop. Each executed trajectory changes the vehicle state and subsequent observations, so small tracking deviations can accumulate across steps. Bounded adjustments, smoothing, and fallback limit their effect on subsequent planning and improve closed-loop execution for both Base (+1.23) and A-Flow (+3.21), while 10.02 of A-Flow's +12.00 advantage already comes before post-processing.

Learned corrections and FCT generate and select the planned trajectory; post-processing adjusts it only when kinematic constraints are violated. NAVSIM and HUGSIM use the exact same rules, fixed in advance, with no additional training or tuning for HUGSIM.

Dear Reviewer ctcX,

Thank you once again for your thoughtful review. We have responded to your comments point by point and included the additional experiments in the revised paper.

The comparison you suggested shows that A-Flow retains about 84% of its advantage over Base on HUGSIM when kinematic post-processing is turned off. Across all 436 closed-loop scenarios, A-Flow led Base by 10.02 HD-Score points with post-processing disabled for both methods, compared with 12.00 points with it enabled for both. On NAVSIM, turning it off reduced A-Flow’s PDMS by only 0.0033 points (Appendix H.4).

We also tested your suggestion to sample 32 matrices at random from the same bank. Retrieving 32 matrices by similarity produced better candidates on average before selection and improved the final NAVSIM score by 1.21 PDMS points over random sampling (Appendix H.2).

With the discussion period closing on November 18, please feel free to let us know if there is anything else you would like us to clarify before then.

We sincerely appreciate your time and feedback.

Best regards, The authors of Submission 44816

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

A-Flow learns trajectory corrections while keeping a pretrained flow planner frozen. The authors optimize scene-specific skew-symmetric matrices using driving rewards, store them in a correction library, and retrieve relevant matrices for new scenes. The resulting candidates share a flow prefix and are evaluated by learned value and safety modules. The paper reports results on NAVSIM, zero-shot closed-loop transfer to HUGSIM, and adaptation to a ReCogDrive-based flow planner.

Critical strengths
  1. Learning reusable corrections without updating the base planner is a useful direction. Applying the matrices only at the final sampling step avoids repeating the full flow computation for each candidate.

  2. The gains over the base planner are substantial: NAVSIM PDMS improves from 85.4 to 90.7, and HUGSIM HD-Score from 24.26 to 37.49. The experiment on a ReCogDrive-based flow planner provides further evidence that the approach can be applied to another planner.

  3. Table 4 compares random, learned unconstrained, and learned skew-symmetric matrices, reporting both oracle and selected scores. These results support a benefit from learning the corrections. The paper also distinguishes the distribution-preservation result for its reference dynamics from the properties of the actual implementation.

Critical weaknesses
  1. A standard multi-sample-and-select baseline is missing. The existing ablations support learned matrices over random matrices, but do not compare against generating 32 independent trajectories from the original flow planner and then applying value and safety selection. Please consider this baseline with matched scorer architectures and training budgets. Reporting both planning quality and inference time would help establish what matrix correction adds beyond ordinary multi-sample selection, since the two generation procedures have different costs.

  2. The scene specificity and transfer limits of the matrices remain unclear. Applying matrices from deliberately mismatched scenes would help reveal when corrections cause negative transfer. On representative OOD scenes, it would be useful to apply each matrix in the learned library to the same base trajectory, before selection or fallback, and report the counts and magnitudes of positive, negative, and negligible changes in the actual score. This would show how broadly corrections transfer and how much the system relies on rejecting harmful ones. If a full evaluation is too costly within the response period, a smaller scene subset and a discussion of the remaining limitations would be reasonable.

  3. There is no completed performance comparison under a matched training budget. The method adds matrix optimization, library filtering, and evaluator training. Appendix F reports computational costs and short-run Flow-GRPO profiling, but does not establish performance at an equal training budget. Please consider comparing against conventional reward fine-tuning or a simple adaptation module under the same reward-query or compute budget, including evaluator training costs.

Questions for authors
  1. How does matrix correction compare with 32 independent samples from the original planner followed by an equally trained selector, in both PDMS and latency?

  2. Before selection and fallback, how often do transferred matrices improve or harm OOD trajectories? Does this change when matrices are retrieved by similarity versus deliberately taken from mismatched scenes?

  3. At a matched reward-query or compute budget, how does the full method compare with conventional fine-tuning or a simpler correction module?

5 comments
AuthorsOpenReview ↗

Candidate quality and generation cost

Thank you for your suggestions. Learned corrections improve candidate quality and transfer to difficult scenes through retrieval and selection. A-Flow also outperforms direct fine-tuning under comparable training budgets.

Candidate quality and generation cost

W1. A standard multi-sample-and-select baseline is missing. The existing ablations support learned matrices over random matrices, but do not compare against generating 32 independent trajectories from the original flow planner and then applying value and safety selection. Please consider this baseline with matched scorer architectures and training budgets. Reporting both planning quality and inference time would help establish what matrix correction adds beyond ordinary multi-sample selection, since the two generation procedures have different costs.

Q1. How does matrix correction compare with 32 independent samples from the original planner followed by an equally trained selector, in both PDMS and latency?

Learned corrections produce better candidates at substantially lower generation cost. We use the same frozen planner and retrain FCT on Independent Flow's candidates with the same architecture and training budget. At K=32K=32, A-Flow32 outperforms Independent Flow32 by 2.49 PDMS points, using approximately 1/32 of its candidate generation FLOPs. Results cover all 12,146 navtest scenes.

MethodCandidatesOracle@KFinal PDMSCandidate generation (GFLOPs)
Independent Flow323291.0387.821,181.17
Independent Flow646491.6088.092,362.23
A-Flow323295.1390.3137.02

Doubling Independent Flow to K=64K=64 still leaves both its oracle and final scores below those of A-Flow32. At B=1B=1 in the original evaluation setup (Table 5), A-Flow32 reduces serial candidate generation time from approximately 1.44 s to 52.49 ms.

The bank reuses correction directions learned from rewards in source scenes. All candidates share the first four flow steps and the frozen network output at the final step, with different corrections applied only at t⋆=0.8t_\star = 0.8. Generating 32 candidates therefore requires just five shared network evaluations; the extra work is mainly terminal matrix operations.

In a separate concurrent benchmark on a single A100 (B=8B=8), end-to-end inference (including input processing, candidate generation, and FCT selection) achieves 27.05 scenes/s for A-Flow32 and 18.64 scenes/s for full A-Flow, corresponding to 36.97 and 53.64 ms/scene. All PDMS scores above use the original evaluation configuration.

AuthorsOpenReview ↗

Correction transfer and scene mismatch

Correction transfer and scene mismatch

W2. The scene specificity and transfer limits of the matrices remain unclear. Applying matrices from deliberately mismatched scenes would help reveal when corrections cause negative transfer. On representative OOD scenes, it would be useful to apply each matrix in the learned library to the same base trajectory, before selection or fallback, and report the counts and magnitudes of positive, negative, and negligible changes in the actual score. This would show how broadly corrections transfer and how much the system relies on rejecting harmful ones. If a full evaluation is too costly within the response period, a smaller scene subset and a discussion of the remaining limitations would be reasonable.

Q2. Before selection and fallback, how often do transferred matrices improve or harm OOD trajectories? Does this change when matrices are retrieved by similarity versus deliberately taken from mismatched scenes?

Similarity-based retrieval favors beneficial corrections, while deliberate mismatches increase harmful ones. Selection then yields substantial gains on difficult OOD scenes. We apply the entire correction bank to the same OOD scenes and compare standard retrieval, random sampling, and deliberate mismatch.

The OOD threshold is the 95th percentile of nearest-neighbor map-descriptor distances in the training set, excluding neighbors from the same recording. This identifies 1,364 scenes. We compare each correction with Base in the same scene, before FCT selection or Base fallback. Beneficial, negligible, and harmful changes correspond to ΔPDMS>1\Delta\mathrm{PDMS} > 1, ∣ΔPDMS∣≤1|\Delta\mathrm{PDMS}| \le 1, and ΔPDMS<−1\Delta\mathrm{PDMS} < -1, respectively.

Candidate source (K=32K=32 except full bank)Total candidatesBeneficial: count (%)Negligible: count (%)Harmful: count (%)
Full bank (12,767 matrices)17,414,1885,535,318 (31.79%)8,291,167 (47.61%)3,587,703 (20.60%)
RRF retrieval43,64818,364 (42.07%)18,851 (43.19%)6,433 (14.74%)
Uniform random sampling43,64813,858 (31.75%)20,817 (47.69%)8,973 (20.56%)
Low-similarity mismatch43,64812,508 (28.66%)21,828 (50.01%)9,312 (21.33%)
Navigation-command mismatch43,64810,515 (24.09%)20,693 (47.41%)12,440 (28.50%)

Standard retrieval more often matches driving intent; the navigation-command mismatch condition forces different source and target commands. Retrieval thus improves the candidate pool before selection.

For each category, we first average ΔPDMS\Delta\mathrm{PDMS} within each scene, then weight all scenes containing that category equally. The mean beneficial/harmful changes are +14.00/−20.43 for the full bank and +12.61/−25.65 for RRF. By rejecting harmful candidates and retaining Base when appropriate, FCT raises mean OOD PDMS from 82.92 to 86.84.

Candidate sourcePDMS after selectionGain over BaseScenes retaining Base
RRF retrieval86.84+3.9221.92%
Uniform random sampling86.04+3.1218.48%
Low-similarity mismatch85.53+2.6120.23%

The group with the lowest Base PDMS sees the largest gain. Grouping scenes by the fraction of beneficial matrices shows that selection can find useful corrections even when they are rare in the bank.

Fraction of beneficial matrices per sceneScenes: count (%)Base PDMSPDMS after FCT selectionNet change
0%400 (29.33%)97.1195.07−2.04
(0%, 10%]100 (7.33%)32.1251.84+19.72
(10%, 50%]445 (32.62%)78.7985.05+6.26
(50%, 100%]419 (30.72%)85.8989.22+3.33
All OOD scenes1,364 (100%)82.9286.84+3.92

In the 0% group, Base averages 97.11 and no matrix improves on Base by more than 1 point; across the remaining 70.67% of scenes, FCT achieves net gains in every group, with 15.47% of all OOD scenes scoring below Base after selection. Retrieval improves the candidate pool; FCT uses predicted value and safety in the target scene to choose among corrections and Base.

AuthorsOpenReview ↗

Comparison with direct Flow-GRPO

Comparison with direct Flow-GRPO

W3. There is no completed performance comparison under a matched training budget. The method adds matrix optimization, library filtering, and evaluator training. Appendix F reports computational costs and short-run Flow-GRPO profiling, but does not establish performance at an equal training budget. Please consider comparing against conventional reward fine-tuning or a simple adaptation module under the same reward-query or compute budget, including evaluator training costs.

Q3. At a matched reward-query or compute budget, how does the full method compare with conventional fine-tuning or a simpler correction module?

A-Flow outperforms direct Flow-GRPO at comparable training cost. Both methods train on four H200 GPUs under a total budget of approximately 88 H200 GPU·h (about 22 wall-clock hours). A-Flow's cost includes 250-step matrix optimization (A250) and FCT training; bank scoring and filtering are listed separately. Full A-Flow scores 4.52 PDMS points higher on the full navtest set.

MethodTraining cost (H200 GPU·h)Bank transfer scoring and filteringFinal PDMS
Flow-GRPO88.01—86.18
A-Flow3288.86 (A250: 80.36 + FCT: 8.50)H200, 17.05 min90.31
Full A-Flow88.86 (A250: 80.36 + FCT: 8.50)H200, 17.05 min90.70

Matrix optimization is bounded by CPU reward scoring rather than GPU compute. A250 caches the frozen prefix and terminal transition terms, leaving mainly CPU-based PDM simulation scoring and 24×2424 \times 24 matrix updates, so average GPU utilization during matrix optimization is 1.47%. Flow-GRPO repeatedly runs five-step trajectory generation and decoder backpropagation, averaging 80.49% GPU utilization.

Flow-GRPO updates multiple sampling steps using a nonsmooth terminal PDMS reward. Credit-assignment variance and cross-scene interference in the shared decoder can limit gains within a finite budget. A-Flow freezes the backbone, optimizes orthogonal correction directions independently per source scene, and handles transfer through a separate selection stage.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful evaluations and constructive feedback. We appreciate their recognition of the value of reusable corrections for frozen planners and the empirical improvements across driving benchmarks. We have addressed their concerns in the individual responses and revised the manuscript accordingly.

A-Flow learns reusable, reward-guided corrections at the frozen planner's final sampling step. Below, we summarize the additional experiments and clarifications, together with the corresponding reviewer comments and manuscript sections.

Controlled Comparisons and Training Cost

  • Independent sampling and selection. With the same frozen planner and equally trained selectors, A-Flow32 achieves 90.31 PDMS, compared with 87.82 for independent Flow32, using approximately 1/32 of the candidate-generation FLOPs. Increasing independent samples to 64 still leaves both oracle and selected scores below A-Flow32. [DSmQ W1/Q1; TYv3 W1/Q1; Appendix H.1, Table S3.]
  • Direct fine-tuning. At comparable training costs of approximately 88 H200 GPU-hours, full A-Flow achieves 90.70 PDMS, versus 86.18 for direct Flow-GRPO. A-Flow's training cost includes matrix optimization and FCT training; bank scoring and filtering are itemized separately. [DSmQ W3/Q3; JATs W3/Q3; Appendix F, Table S2.]

Correction Transfer and Component Contributions

  • Retrieval and target-scene selection. Scene-based retrieval improves final PDMS by 1.21 points over random sampling from the same learned bank. Direct Rank-1 correction without selection underperforms Base, supporting the importance of target-scene evaluation. [ctcX W1/Q1; TYv3 W1; Appendix H.2, Tables S4–S5.]
  • Transfer to OOD scenes. Exhaustive bank evaluation on 1,364 OOD scenes shows that retrieval increases the proportion of beneficial corrections, whereas deliberate scene mismatches increase harmful transfers. With retrieval and FCT selection, mean OOD PDMS improves from 82.92 to 86.84. [DSmQ W2/Q2; Appendix I, Tables S8–S10.]
  • Kinematic post-processing. On HUGSIM, A-Flow retains a 10.02-point HD-Score advantage over Base when post-processing is disabled for both methods. We report results with and without post-processing on both NAVSIM and HUGSIM. [ctcX W2/Q2; Appendix H.4, Table S7.]

Inference Cost and Performance Trade-off

  • Performance and processing time. In the original evaluation configuration (B=1B=1), single-branch Value+Safety selection retains approximately 93% of full A-Flow's PDMS gain over Base, while reducing three-stage processing time from 241.33 to 125.02 ms. [TYv3 W2/Q2; Section 5.3, Table 5.]
  • Concurrent inference. In a separate benchmark on one A100 80GB with eight processes and B=8B=8 per process, the single-branch configuration reduces amortized time from 53.64 to 36.97 ms/scene, a 31.1% reduction. Cumulative measurements quantify the additional costs of scale adjustment and the second retrieval branch. [DSmQ W1/Q1; TYv3 W2/Q2; Section 5.3, Table 6; Appendix J.]

Method Clarification and Presentation

  • Role of the correction matrix. Section 4 now explains the role of the matrix before introducing its theoretical motivation. Implementation settings appear in Section 5, with detailed conditioning, optimization, and bank-construction procedures in the appendices. [JATs W1; Sections 4–5; Appendices A, C.2, and D.1.]
  • Parameter and storage accounting. We distinguish per-scene optimized parameters, full and deployed bank storage, and the matrices retrieved for each query. [JATs Q2; Appendix F, Table S1.]

Thank you again for your time and consideration.

Area_Chair_kaZeOpenReview ↗

Reviewers, please help to join the discussion.

Dear reviewers,

The authors have uploaded the rebuttal. Please help to join the discussion and the further process.

Thanks!

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.