acceptodds
Under review as a conference paper at ICLR 2027

Simulation-Free Flow Map Alignment for High-Fidelity Any-Step Constrained Generation

10 views

Abstract

Flow-based generative models for science and engineering should satisfy physical and operational constraints while preserving distributional fidelity. Inference-time correction improves constraint satisfaction but adds per-sample cost and can distort the generated distribution. Training-time alignment amortizes this cost, but propagating terminal feedback to the local dynamics typically requires costly ODE simulation. Flow map models avoid simulation by evaluating terminal losses directly on learned finite-time transitions. Yet endpoint training alone neither yields the local-velocity update implied by the ODE objective nor ensures that alignment persists when shorter transitions are composed. We introduce simulation-free terminal alignment for any-step flow map generation. A map-based adjoint uses learned intermediate states and the sensitivity of the remaining-time map to transport terminal feedback to the local velocity, approximating the gradient of the ODE terminal objective. Combined with endpoint training and map–velocity consistency, this update aligns a single model across sampling schedules. Our analysis characterizes the map-adjoint approximation error and shows how velocity and consistency errors affect distributional fidelity and cross-schedule agreement. Experiments on constrained distributions, robot trajectories, and physical fields show that the aligned model retains constraint improvements across multi-step schedules while maintaining competitive distributional quality.

Primary area:
generative models
Keywords:
generative models, constrained generation, training-time alignment, adjoint method, scientific machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fK3uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript is about how to make the generation of flow-based generative models satisfy constraints while preserving distributional fidelity. This work learns the terminal constraint into the local dynamics of a flow map without any rollout, and keeps this alignment when the number of sampling steps changes, by proposing a simulation-free terminal alignment for any-step flow map generation.

The interesting and novel part is the map-based adjoint. The classical Pontryagin adjoint needs the intermediate state xtx_t, the endpoint x1x_1, and the Jacobian of the remaining-time flow; this work replaces all of them with the learned map: x^t=T(z;0,t)\hat{x}_t = T(z; 0, t), y^=T(x^t;t,1)\hat{y} = T(\hat{x}_t; t, 1), then one VJP gives a^t\hat{a}_t, which is injected into autograd through a zero-forward-value surrogate loss. The cost per update is fixed and independent of the number of ODE steps.

Critical strengths
  1. The problem is well framed and well illustrated in Table 1, i.e., training-time constraint alignment, any-step generation, and no ODE rollout during training are required at the same time, which has not been addressed before.
  2. The map-based adjoint is simple and cheap. It replaces the intermediate state, endpoint, and remaining-time Jacobian of the Pontryagin adjoint with the learned map and one VJP, so it can be plugged into any flow model with fixed cost per update.
  3. The ablation is well designed.
Critical weaknesses

All six benchmarks are either 2D toy distributions or small MLPs on POD latents of at most 144 dimensions. So I was wondering how the method performs in high-dimensional settings.

Questions for authors
  1. It seems that VFM and Direct map give almost same numbers everywhere. Could the authors explain why the noise adapter of VFM seems to have no effect here?
  2. Ours, VFM, and Direct map start from the FMM–LSD parent, while ODE-Adjoint, Flow-GRPO, and FM-RE start from FM (Euler), which is necessarily poor at small KK. Could the authors add an ODE-Adjoint baseline on the same FMM–LSD parent?
  3. Compared with FMRG, which uses the Jacobian of the remaining-time map for inference-time guidance, this work moves the same quantity from inference time to training time. Could the authors clarify the increment over FMRG?
  4. Could the authors discuss the performance in high-dimensional settings, or the limitation there?
6 comments
AuthorsOpenReview ↗

Response (Part 1)

Thank you for the constructive review and for recognizing the value of a simple, efficient map-based adjoint for training-time alignment and the ablations supporting its role. Below, we respond point by point to your concerns, with additional comparisons that clarify baseline fairness, the contribution relative to prior work, and performance at larger latent dimensions.


C1

(W1) All six benchmarks are either 2D toy distributions or small MLPs on POD latents of at most 144 dimensions. [...] how the method performs in high-dimensional settings.

(Q4) [...] discuss the performance in high-dimensional settings, or the limitation there?

Response.

The submission includes a 512-dimensional latent generator for Dynamic Stall, and we now extend the evaluation to learned dimensions 1,024 and 2,048. At both larger dimensions, Ours improves violation and SWD over same-parent Direct controls across all three training seeds and tested step counts.

  • Submitted scale. The submission uses learned dimensions 144, 128 and 512 for SMD9, Stokes and Dynamic Stall (Table 2). Dynamic Stall decodes to 98,304 physical coordinates; its submitted latent generator is a width-128, depth-3 MLP. Learned and decoded dimensions are distinct.

  • New larger-state comparison. Across three training seeds, Direct and Ours start from the same FMM parent within each rank, with the same backbone and training budget. Paired tests cover 12 conditions and K=1,2,4,8,16,32,64,128K=1,2,4,8,16,32,64,128. The codecs are fitted on the same training data.

    POD rankDirect violation (10−610^{-6})Ours violation (10−610^{-6})Direct SWDOurs SWD
    5127.876±0.4317.876\pm0.4317.287±0.3207.287\pm0.3200.03562±0.000930.03562\pm0.000930.03515±0.000890.03515\pm0.00089
    102410.018±0.24710.018\pm0.2479.064±0.2579.064\pm0.2570.04159±0.000640.04159\pm0.000640.04137±0.000620.04137\pm0.00062
    204811.313±0.58211.313\pm0.5829.915±0.3549.915\pm0.3540.04883±0.000970.04883\pm0.000970.04850±0.000730.04850\pm0.00073

    Entries show mean ±\pm sample SD across three training seeds, after averaging the 12 conditions and eight step counts within each seed. At each larger rank, Ours improves both metrics over Direct in all 24 seed–step comparisons, reducing mean violation by 9.5% at rank 1024 and 12.4% at rank 2048. The SWD improvements are small. Thus the within-rank benefit over Direct persists, while increasing rank does not improve the absolute generation metrics.

  • Computation and memory at larger dimensions. Direct and Ours have identical trainable parameter counts at each rank; the frozen POD basis is stored separately. On one H200, the following measurements use the final checkpoints from one training run per method and batch size 64. Paired entries are Direct/Ours.

    RankParametersCodec MiBFwd+bwd time (ms)Peak MiB
    512281,472192.45.08/14.40403/543
    1024413,056384.45.10/14.12603/744
    2048676,224768.45.08/14.731003/1146

    Times are medians over repeated measurements, excluding optimizer updates and data preparation. Peak allocated memory includes the model, codec and runtime buffers. The frozen basis accounts for much of the memory growth. Avoiding an ODE rollout removes dependence on its integration length, while evaluating larger maps and differentiating through their decoders still carries a cost.

These results show that alignment improves violation and SWD over Direct at larger POD dimensions, with increased memory use as the model and decoder grow.

Revised Appendix D.7 now reports the reconstruction comparison, three-seed generation results, tolerance attainment, and measured model, codec and computation costs (Tables 29–31). Revised Sec. 7 now discusses the measured state-size scaling and the role of model capacity and decoder storage.

AuthorsOpenReview ↗

Response (Part 2)

C2

(Q1) [...] VFM and Direct map give almost same numbers everywhere. [...] explain why the noise adapter of VFM seems to have no effect here?

Response.

VFM and Direct map both train the flow map through the terminal penalty. In our evaluated VFM models, the additional source adapter contributes little to the reported metrics. The inspected Six Disks checkpoint learns a source distribution very close to the base Gaussian. Replacing the adapter by the identity while keeping the trained map fixed has little effect on violation across all four tested tasks, with a modest SWD benefit from the adapter on Six Disks.

  • Joint training and learned source distribution. Our submitted VFM implementation applies terminal feedback to both the adapter and map, with KL regularization of the adapted source toward the Gaussian prior (Appendix C.2). We retain the KL coefficient of 11 used in the public VFM implementation (Mammadov et al., 2026). For the inspected Six Disks checkpoint, the adapted Gaussian has a mean-vector norm of 0.00190.0019, mean standard deviation 1.00181.0018, and KL divergence 1.11×10−51.11\times10^{-5} from the base Gaussian. The learned source therefore differs little from the original source distribution at this checkpoint.

  • Effect of replacing the adapter. We hold the trained map and base draws fixed and replace only the learned source adapter by the identity at deployment. Across eight uniform step counts, mean violation changes from 0.00309790.0030979 to 0.00310050.0031005 on Six Disks, from 0.01414990.0141499 to 0.01415160.0141516 on SMD9, and from 0.01170850.0117085 to 0.01171960.0117196 on Stokes. Dynamic Stall remains approximately 7.8×10−67.8\times10^{-6}. SWD changes from 0.051190.05119 to 0.052480.05248 on Six Disks, 0.03571250.0357125 to 0.03571550.0357155 on SMD9, and 0.13142520.1314252 to 0.13136080.1313608 on Stokes; Dynamic Stall changes by less than 10−610^{-6}. Six Disks uses one training seed; the other tasks use three, with means over test conditions and step counts, using independent source draws. This replacement test confirms that the learned adapter adds little to the measured violation reduction beyond the trained map. Its small additional contribution helps explain why VFM and Direct map have similar scores here.

The comparison with and without the adapter is now reported in revised Appendix D.3, Table 17.

AuthorsOpenReview ↗

Response (Part 3)

C3

(Q2) Ours, VFM, and Direct map start from the FMM–LSD parent, while ODE-Adjoint, Flow-GRPO, and FM-RE start from FM (Euler), which is necessarily poor at small K. [...] add an ODE-Adjoint baseline on the same FMM–LSD parent?

Response.

We added an ODE-feedback control with the same FMM–LSD parent, loss components and learned-map sampler as Ours, using eight Euler steps for training feedback to match the submitted ODE-Adjoint rollout budget. The new comparison shows a trade-off: Ours achieves lower violation at every tested step count, while ODE feedback achieves lower SWD and has a longer observed alignment-training time.

  • Controlled comparison. The submitted ODE-Adjoint starts from FM (Table 5, p. 24). The new control and Ours start from the same FMM–LSD checkpoint and use the same network, FM/direct/LSD terms, feedback coefficient of 1, optimizer and 10,000-update budget. Only the terminal-feedback estimator changes: the control differentiates through an eight-step Euler rollout instead of using the map adjoint. At deployment, both models compose learned maps. The sampling step count KK below therefore refers to map evaluations, not the eight Euler steps used during training.

  • New paired results. The following Six Disks comparison uses one alignment seed and 4,096 paired validation draws per schedule at K=1,4,16,64K=1,4,16,64. Violation is reported in units of 10−310^{-3}.

    Steps KKOurs violationODE feedback violationOurs SWDODE feedback SWD
    11.6411.8090.103220.08487
    40.4291.6860.112490.08407
    160.3620.6790.111460.07900
    640.3700.5770.110880.07846

    Averaged over these schedules, Ours reduces violation from 1.1881.188 to 0.7010.701, while SWD increases from 0.081600.08160 to 0.109510.10951. Thus the constraint benefit persists with a matched parent and sampler, with a trade-off in distributional fidelity.

  • Training cost. The recorded 10,000-update alignment runs take 5.415.41 min for ODE feedback and 2.542.54 min for Ours, a ratio of 2.13×2.13\times, excluding the common parent pretraining. These are observed wall times under shared GPU use. The map adjoint supplies terminal feedback without the sequential integration required by the ODE control.

Revised Appendix D.6 reports the matched comparison and alignment times (Tables 25–26). Appendix D.8 gives the separate training-stage costs.

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

This paper aligns flow-map generative models with a terminal constraint penalty without ODE rollouts. A shared network represents a local velocity and finite-time maps. Training combines flow matching, a one-step endpoint penalty, a map-based adjoint that transports terminal gradients through a learned map Jacobian, and map-velocity consistency. The authors give error bounds and evaluate one frozen model at 1, 4, 16, and 64 uniform sampling steps on six constrained-generation benchmarks.

Critical strengths
  1. The method targets a concrete bottleneck: terminal feedback normally requires expensive rollouts, while endpoint fine-tuning may not transfer when shorter maps are composed.
  2. The paper supports its central claim with an adjoint error analysis, cross-schedule evaluation, and ablations of the map-adjoint and consistency terms.
Critical weaknesses
  1. The central approximation relies on a learned map Jacobian that the training objective does not directly constrain. The error bound contains a Jacobian-error term, and the authors' diagnostic reports opposite gradients for 17% of nonzero pairs on Six Disks and 22% on Checkerboard. The diagnostic covers frozen pre-alignment toy models rather than the parameter gradient used during training, leaving the reliability of the proposed update insufficiently established.

  2. The experiments show reduced violations, not physical feasibility. Every evaluated Dynamic Stall sample exceeds the stated residual tolerance, and the frozen POD codec also exceeds it on all 1,024 held-out validation reconstructions. Since both generator and representation violate the tolerance, the paper cannot separate model error from decoder error or validate the high-fidelity physical claim.

  3. Evidence for any-step behavior and generality is narrow. Evaluations use only uniform schedules with K in {1,4,16,64}; higher-dimensional tasks use small frozen POD latents. Results are mixed across benchmarks, with several quality-constraint trade-offs, and no code or artifact entry is provided.

Questions for authors
  1. Can you report post-alignment Jacobian errors, adjoint cosine/parameter-gradient agreement, and missed or opposite gradients across all benchmarks?
  2. Can you decompose Dynamic Stall residual error into generator versus POD/decoder error, and test less compressed representations?
  3. Do gains persist under nonuniform schedules and K>64? Were the loss weights selected on validation data, and will code and full-precision records be released?
6 comments
AuthorsOpenReview ↗

Response (Part 1)

Thank you for the careful review and for recognizing the importance of efficient terminal feedback that remains useful under composed sampling, together with the analysis and ablations supporting our design. Below, we address your concerns point by point, distinguishing what the submission establishes from the additional evidence on update reliability, physical residuals, and sampling schedules.


C1

(W1) The central approximation relies on a learned map Jacobian that the training objective does not directly constrain. The error bound contains a Jacobian-error term, [...] The diagnostic covers frozen pre-alignment toy models rather than the parameter gradient used during training, leaving the reliability of the proposed update insufficiently established.

(Q1) Can you report post-alignment Jacobian errors, adjoint cosine/parameter-gradient agreement, and missed or opposite gradients across all benchmarks?

Response.

We agree that the key question is whether map-adjoint feedback supplies a penalty descent direction after alignment updates. We now test this through parameter-gradient comparisons at initialization, during Six Disks alignment, and on SMD9, Stokes and Dynamic Stall. The aligned cross-task checkpoints all have positive cosine at the tested Euler resolutions.

  • Why compare parameter gradients? Jacobian operator error measures derivative discrepancies across state-space directions. The training feedback instead uses the direction selected by the terminal penalty, maps it to network parameters, and aggregates contributions across samples and times. Comparing the resulting parameter gradients directly tests whether this feedback decreases the penalty. LSD does not directly constrain Jacobian error, so that term remains explicit in submitted Eq. (18).

    Both computations use the same network checkpoint, terminal penalty and 8,192 source samples. The map-feedback gradient averages the adjoint training contribution over 16 time draws per source. The Euler gradient differentiates the terminal penalty through an 8-, 16-, 32- or 64-step rollout. We include all trainable network parameters and all samples, including zero feedback, and compute cosine after aggregating the gradients. This tests the net effect of zero and opposing feedback on the parameter direction. An opposite direction here means a negative aggregate cosine, which is distinct from the submitted per-sample state-adjoint counts. For the Euler terminal objective JNJ_N, gEuler,N=∇θJNg_{{\rm Euler},N}=\nabla_\theta J_N and the map-feedback mean gmapg_{\rm map} satisfy ddηJN(θ−ηgmap)∣η=0=−gEuler,N⊤gmap\left.\frac{\mathrm{d}}{\mathrm{d}\eta}J_N(\theta-\eta g_{\mathrm{map}})\right\rvert_{\eta=0}=-g_{\mathrm{Euler},N}^{\top}g_{\mathrm{map}}. Positive cosine thus identifies local descent for that discrete terminal objective.

  • Evidence during alignment and across tasks. At the Six Disks FMM–LSD initialization, the cosines are 0.7540.754, 0.3930.393, 0.1260.126 and 0.6080.608 for Euler-8/16/32/64. We extend this comparison beyond the frozen initialization as follows:

    • Training checkpoints. At updates 2,500, 5,000 and 7,500, all four cosines are positive, ranging from 0.0940.094 to 0.4220.422. At update 10,000, they are −0.323-0.323, 0.0720.072, 0.0090.009 and 0.3270.327 for Euler-8/16/32/64. Thus agreement persists in most measured comparisons. At the final checkpoint, the cosine is negative against the coarse eight-step Euler approximation and positive against the finer Euler-16/32/64 approximations.

    • Cross-task checkpoints. For aligned models, the three-seed mean cosines for Euler-8/16/32/64 are 0.9420.942, 0.9440.944, 0.9440.944, 0.9450.945 on SMD9, 0.4530.453, 0.4970.497, 0.5180.518, 0.5270.527 on Stokes, and 0.9610.961, 0.9630.963, 0.9610.961, 0.9600.960 on Dynamic Stall. Every individual aligned checkpoint has positive cosine at all four resolutions. Each comparison aggregates four validation conditions using shared source samples. The revised appendix also reports the FMM–LSD checkpoints and the range across seeds.

    Thus the cross-task results support using map feedback as a local penalty descent direction. The training checkpoints show that this direction can change with alignment and Euler resolution.

Revised Sec. 4.2 explains the parameter-space comparison. Appendix D.2, Tables 14–15, reports the training and cross-task results.

AuthorsOpenReview ↗

Response (Part 2)

C2

(W2) The experiments show reduced violations, not physical feasibility. Every evaluated Dynamic Stall sample exceeds the stated residual tolerance, and the frozen POD codec also exceeds it on all 1,024 held-out validation reconstructions. Since both generator and representation violate the tolerance, the paper cannot separate model error from decoder error or validate the high-fidelity physical claim.

Response.

The submitted uncorrected Dynamic Stall outputs do not meet the prescribed tolerance: soft alignment reduces residuals without guaranteeing hard feasibility. Our new comparison quantifies the residual introduced by POD reconstruction alongside the larger generated-field residual. New physical-space correction then meets the tolerance for the two evaluated constraints on every tested sample, with a small SWD increase.

  • POD reconstruction and generation error. The submission already reports that all 1,024 held-out validation reconstructions exceed tolerance, with mean excess residual 1.84×10−71.84\times10^{-7} (Appendix D.1.6). Our new comparison uses raw reference fields, their POD-512 reconstructions, and generated fields on the same 12 test conditions across three training seeds. The normalized-field reconstruction MSE is 3.00×10−63.00\times10^{-6}. Mean physical residuals are 7.90×10−107.90\times10^{-10} for raw references and 1.71×10−71.71\times10^{-7} for their reconstructions; at K=16K=16, they are 8.72×10−68.72\times10^{-6} for the FMM–LSD parent and 7.71×10−67.71\times10^{-6} for Ours. Thus POD reconstruction introduces residual, while alignment reduces the substantially larger generated-field residual within the same representation. These measurements quantify the reconstruction effect, rather than give an additive decomposition of generated-field error.

  • Physical-space correction and distributional cost. The submitted correction uses finite Lagrangian iterations in latent space and does not reach the tolerance. In the new experiment, we instead correct pressure and wall shear directly using their algebraic relations to the other fields. This enforces the two evaluated constraints up to numerical precision and can leave the POD decoder’s range.

    Using existing FMM–LSD, Direct and Ours checkpoints across three seeds and K=1,4,16,64K=1,4,16,64, every evaluated sample meets the original tolerance 2.5×10−92.5\times10^{-9} after correction. For Ours at K=16K=16, mean physical residual decreases from 7.41×10−67.41\times10^{-6} to 2.48×10−172.48\times10^{-17}, attainment rises from 0% to 100%, and SWD changes from 0.0352040.035204 to 0.0353130.035313. These are means over three seeds and 12 test conditions, using independent draws paired before and after correction, separate from the reconstruction comparison above. Across method–seed–schedule comparisons, the relative SWD increase is 0.29%–0.38%. This correction complements soft alignment by enforcing the evaluated constraints with a small measured distributional cost.

Revised Sec. 1 states the soft-penalty objective, and Sec. 7 distinguishes alignment from inference-time enforcement. Revised Appendix D.7 reports the reconstruction comparison and Table 28 gives the generation results. Table 32 now reports the physical-space correction comparison in the same appendix.

AuthorsOpenReview ↗

Response (Part 3)

C3

(W3) [...] Evaluations use only uniform schedules with K in {1,4,16,64}; higher-dimensional tasks use small frozen POD latents. Results are mixed across benchmarks, with several quality-constraint trade-offs, and no code or artifact entry is provided.

Response.

Our any-step evaluation tests whether one trained model remains effective across sampling budgets without retraining. The added step counts and nonuniform schedules extend this test beyond the submitted uniform grids. On Six Disks, the violation-reduction benefit persists across the added budgets and partitions. On the additional tasks, Ours maintains similar violation across the tested grids. We also extend the learned-state comparison to POD dimensions 1,024 and 2,048.

  • Additional step counts. The submission uses uniform K=1,4,16,64K=1,4,16,64 to compare one-step, few-step and longer compositions. On Six Disks, we now add intermediate budgets K=2,8,32K=2,8,32 and a larger budget K=128K=128, keeping one trained checkpoint per method fixed. Ours has lower mean violation than Direct map and VFM at all four additional budgets, with higher SWD. At K=128K=128, mean violation for Ours, Direct map and VFM is 0.3990.399, 1.7161.716, and 1.7111.711 in units of 10−310^{-3}, while SWD is 0.0870.087, 0.0610.061, and 0.0610.061. These results use one training seed and 4,096 paired validation draws per schedule.

    For SMD9, Stokes, and Dynamic Stall, we evaluated eight uniform budgets, K=1,2,4,8,16,32,64,128K=1,2,4,8,16,32,64,128, using fixed trained checkpoints from three training seeds and fresh source samples. Equal-weight means over seeds, test conditions, and these step counts give Ours/VFM violation of 0.01037/0.014150.01037/0.01415, 0.00863/0.011710.00863/0.01171, and 7.3×10−6/7.8×10−67.3\times10^{-6}/7.8\times10^{-6}, respectively.

    For the same tasks, the corresponding mean SWD values are 0.03070/0.035710.03070/0.03571, 0.13666/0.131430.13666/0.13143, and 0.03515/0.035390.03515/0.03539. Thus SMD9 improves both averages, Stokes retains a fidelity–violation trade-off, and Dynamic Stall shows a small improvement.

  • Nonuniform partitions. On Six Disks, we evaluated 15 nonuniform partitions at K=4,16,64K=4,16,64 using the same fixed checkpoints and paired validation draws as above. Ours has lower mean violation than Direct map and VFM on all 15 partitions, with higher SWD.

    On SMD9, Stokes, and Dynamic Stall, we compare Ours across uniform and nonuniform grids. Three fixed nonuniform K=16K=16 grids give mean violation ranges 0.008770.00877–0.009670.00967 on SMD9, 0.008260.00826–0.008880.00888 on Stokes, and approximately 7.37.3–7.4×10−67.4\times10^{-6} on Dynamic Stall, compared with uniform-grid values 0.009190.00919, 0.008730.00873, and 7.40×10−67.40\times10^{-6}. This within-model comparison shows that Ours maintains similar violation across the tested grids.

  • Larger learned states. The POD-1024 and POD-2048 comparisons retain the benefit of alignment over Direct map under the same backbone and training budget. Increasing rank alone does not improve absolute generation quality. C4 gives the reconstruction measurements and three-seed generation results.

  • Code and records. The final version will include a GitHub link to the complete code, data and trained models; C5 details the release contents.

The uniform and nonuniform comparisons are now in revised Appendices D.4–D.5, Fig. 4 and Tables 18–24, with their sampling settings and task-dependent quality–constraint interpretation. Revised Appendix D.7 reports the larger-POD comparisons, and Appendix C.4 states the release plan.

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

This paper studies how to enforce terminal constraints in flow-based generative models while preserving distributional fidelity and keeping generation efficient across different sampling budgets.

Flow-based models learn a velocity field that transports noise to data via an ODE. In scientific and engineering applications, generated samples must also satisfy physical constraints such as feasible sets, collision avoidance, or PDE residuals. Inference-time correction via guidance or projection improves constraint satisfaction but adds per-sample cost and can distort the distribution. Training-time alignment amortizes this cost by penalizing terminal violations during training, but the terminal loss is defined on the endpoint of the ODE flow, so its gradient with respect to the local velocity requires either backpropagation through a numerical trajectory or trajectory-based policy gradients, both requiring rollouts.

Flow map models, including Flow Map Matching, Shortcut Models and MeanFlow, avoid simulation by learning finite-time transitions directly. A shared network predicts the normalized displacement over an interval and its evaluation at zero interval defines the local velocity. Training uses a Lagrangian self-distillation loss that penalizes the residual between the time derivative of the map and the velocity evaluated at the predicted state. After training, the map can be sampled in one step or composed over a schedule of shorter transitions, enabling any-step generation with one model.

The authors argue that directly training the one-step map output with the terminal penalty does not supply the local-velocity update implied by the ODE terminal objective and does not guarantee that the improvement persists when the map is composed. Their contribution is a simulation-free terminal alignment for any-step flow map generation.

The method has two components. First, a map-based adjoint approximates the continuous adjoint identity from optimal control. The exact adjoint expresses the gradient of the terminal loss as local velocity updates along the trajectory weighted by a backward-propagated terminal covector involving the Jacobian of the remaining-time flow. The paper replaces the intermediate state with a learned prefix map and the remaining-time Jacobian with the Jacobian of the learned suffix map. A zero-forward-value surrogate is constructed so that automatic differentiation with stop-gradient yields the approximate local-velocity update without ODE rollout. Second, direct endpoint training penalizes the one-step map output and velocity-map consistency aligns finite transitions with the updated dynamics.

The analysis characterizes the map-adjoint error as a sum of an LSD-residual term controlling state and endpoint accuracy and a Jacobian error term controlling gradient transport, using Gronwall's inequality. It then proves distribution error and cross-schedule consistency bounds: Wasserstein-2 distance to the data distribution is bounded by excess flow-matching risk plus square-root of LSD residual energy along the deployed composition, and terminal-loss difference between two schedules is bounded by the sum of their residual energies. Extensions to quadratic penalties and latent decoder settings are given in appendix.

Experiments evaluate whether alignment learned once retains value across schedules K=1,4,16,64 without retraining. Benchmarks include Six Disks and Checkerboard constrained distributions, SMD6/SMD9 conditioned multi-robot trajectories, and Stokes/Dynamic Stall physical fields, with high-dimensional tasks trained in POD latent space with frozen encoder/decoder and decoded before evaluation. Baselines include unaligned FMM-LSD, VFM joint map and source adapter, FM-RE randomized late-stage flow, ODE-Adjoint backprop through trajectory, Flow-GRPO policy gradient, FMRG-J inference-time map Jacobian guidance, and FM+Proj projection. Metrics are sliced Wasserstein distance versus mean distance outside feasible set, collision penetration, and mean excess physical residual. Training cost and sampling cost are reported on H200.

Results show the aligned model lowers mean terminal metric at every K versus the parent, with modest SWD increase at K=1 and better trade-off at composed schedules versus VFM. Direct map ablation is competitive at K=1 but worse at composed K, supporting the adjoint. Consistency off raises SWD sharply. Training time is about 9-12ms/update versus 22-29ms for ODE-Adjoint and 10-11ms for Flow-GRPO. Inference-time correction can drive violation lower but with larger SWD and schedule instability.

Critical strengths
  1. Distinguishes where constraint information enters, how terminal feedback becomes a parameter update, and how finite transitions compose. Any-step alignment is a well-motivated gap not covered by prior flow-map alignment works.

  2. Using learned intermediate states and remaining-time map Jacobian to approximate the ODE adjoint avoids rollout with fixed number of map evaluations and VJPs. Zero-forward surrogate to inject gradient via autodiff is technically neat.

  3. Coupling adjoint velocity update, direct endpoint loss, and velocity-map consistency addresses both local dynamics and finite transitions, unlike endpoint-only fine-tuning.

  4. Bounds separate excess FM risk and LSD residual energy along actual compositions, explaining why consistency matters for any-step agreement.

Critical weaknesses
  1. LSD penalizes residual but not Jacobian operator norm error. The map-adjoint error bound explicitly contains Jacobian error term not optimized. Empirical support is median cosine 0.994 on Six Disks and 0.972 on Checkerboard against numerical ODE adjoint on frozen pre-alignment flow maps, when both nonzero. This diagnostic does not measure parameter gradient used during training nor does it hold after alignment updates. No result shows Jacobian error decreases with training.

  2. Theorem 2 uses exponential of Lipschitz integral where Lipschitz constant is of learned velocity, and a related constant involving double exponential. For typical neural velocities, this exponential can be huge. Constants for terminal penalty Lipschitz are not quantified for smoothed penalty with delta 0.01. No numerical evaluation of bound is given. Reference assumption that population flow-matching minimizer realizes exact data path requires exact marginal matching by population minimizer, idealized.

  3. Direct map baseline retains flow matching and consistency but backprops terminal loss only through one-step map. Because map and velocity share network, direct loss still changes velocity. Comparison isolates added explicit adjoint update, not direct vs adjoint alone. Lambda sweep only on Six Disks validation, three seeds, no search for violation and consistency weights. Removing consistency lowers violation at small K but raises SWD 0.06 to 0.17 at K=1, suggesting consistency is fidelity regularizer, not violation penalty, but trade-off not disentangled.

  4. Any-step evaluated only at K=1,4,16,64, not arbitrary partitions. Proposition holds for any fixed partition but experiments do not test non-uniform or K=2,8,32. No continuous-time consistency error reported.

  5. Method assumes differentiable penalty, smooth surrogate introduces bias. Extension to black-box rewards suggested as zeroth-order estimation, not implemented. Cost comparison reports optimizer update time but not memory or total training time including pretraining of parent. Sampling cost logarithmic axis hides constant factors.

Questions for authors
  1. Can you bound Jacobian error by LSD energy or regularize it? Would adding Jacobian consistency improve cosine similarity during alignment, not just on frozen maps?

  2. What are empirical values of Lambda and C constants for final checkpoints? Is exponential on order of tens, thousands, or larger? If large, how to interpret Theorem 2?

  3. Since network is shared, what fraction of velocity update in Ours comes from direct endpoint loss versus adjoint? Can you log norms during training?

  4. Why does one-step SWD degrade while composed SWD improves? Does direct endpoint training overfit to full interval, and does consistency mitigate it? Provide residual energies for each K for Ours vs Direct vs VFM.

  5. Report full statistics over seeds for cross-schedule figure, including variance, and statistical test for violation reduction at each K. Why do many means fall outside plot range?

  6. For Dynamic Stall, residual remains above tolerance — does constraint alignment fail for this physics task? What residual would be achieved with exact projection at inference, and what is SWD cost?

  7. Test any-step beyond powers of two: random partitions and K=2,8,32. Does cross-schedule bound predict observed loss differences?

  8. POD setting: quantify decoder reconstruction error on ground truth and its contribution to reported physical residual. Does gradient through frozen decoder cause latent collapse?

10 comments
AuthorsOpenReview ↗

Response (Part 1)

Thank you for the thoughtful review and for recognizing the motivation for any-step alignment and the way our method connects terminal feedback to both local dynamics and finite-time transitions. Below, we respond point by point to your concerns, clarifying the analysis and presenting additional evidence on the proposed update and its behavior across sampling schedules.


C1

(W1) LSD penalizes residual but not Jacobian operator norm error. The map-adjoint error bound explicitly contains Jacobian error term not optimized. Empirical support is [...] frozen pre-alignment flow maps, when both nonzero. This diagnostic does not measure parameter gradient used during training nor does it hold after alignment updates. No result shows Jacobian error decreases with training.

Response.

We agree that the key question is whether map-adjoint feedback supplies a penalty descent direction after alignment updates. We now test this through parameter-gradient comparisons at initialization, during Six Disks alignment, and on SMD9, Stokes and Dynamic Stall. The aligned cross-task checkpoints all have positive cosine at the tested Euler resolutions.

  • Why compare parameter gradients? Jacobian operator error measures derivative discrepancies across state-space directions. The training feedback instead uses the direction selected by the terminal penalty, maps it to network parameters, and aggregates contributions across samples and times. Comparing the resulting parameter gradients directly tests whether this feedback decreases the penalty. LSD does not directly constrain Jacobian error, so that term remains explicit in submitted Eq. (18).

    Both computations use the same network checkpoint, terminal penalty and 8,192 source samples. The map-feedback gradient averages the adjoint training contribution over 16 time draws per source. The Euler gradient differentiates the terminal penalty through an 8-, 16-, 32- or 64-step rollout. We include all trainable network parameters and all samples, including zero feedback, and compute cosine after aggregating the gradients. For the Euler terminal objective JNJ_N, gEuler,N=∇θJNg_{{\rm Euler},N}=\nabla_\theta J_N and the map-feedback gradient gmapg_{\mathrm{map}} satisfy ddηJN(θ−ηgmap)∣η=0=−gEuler,N⊤gmap\left.\frac{\mathrm{d}}{\mathrm{d}\eta}J_N(\theta-\eta g_{\mathrm{map}})\right\rvert_{\eta=0}=-g_{\mathrm{Euler},N}^{\top}g_{\mathrm{map}}. Positive cosine thus identifies local descent for that discrete terminal objective.

  • Evidence during alignment and across tasks. At the Six Disks FMM–LSD initialization, the cosines are 0.7540.754, 0.3930.393, 0.1260.126 and 0.6080.608 for Euler-8/16/32/64. We extend this comparison beyond the frozen initialization as follows:

    • Training checkpoints. At updates 2,500, 5,000 and 7,500, all four cosines are positive, ranging from 0.0940.094 to 0.4220.422. At update 10,000, they are −0.323-0.323, 0.0720.072, 0.0090.009 and 0.3270.327 for Euler-8/16/32/64. Thus agreement persists in most measured comparisons. At the final checkpoint, the cosine is negative against the coarse eight-step Euler approximation and positive against the finer Euler-16/32/64 approximations.

    • Cross-task checkpoints. For aligned models, the three-seed mean cosines for Euler-8/16/32/64 are 0.9420.942, 0.9440.944, 0.9440.944, 0.9450.945 on SMD9, 0.4530.453, 0.4970.497, 0.5180.518, 0.5270.527 on Stokes, and 0.9610.961, 0.9630.963, 0.9610.961, 0.9600.960 on Dynamic Stall. Every individual aligned checkpoint has positive cosine at all four resolutions. Each comparison aggregates four validation conditions using shared source samples. The revised appendix also reports the FMM–LSD checkpoints and the range across seeds.

    Thus the cross-task results support using map feedback as a local penalty descent direction. The training checkpoints show that this direction can change with alignment and Euler resolution.

C6 discusses the distinction between residual and Jacobian error and our early Jacobian-regularization experiments.

Revised Sec. 4.2 explains the parameter-space comparison. Appendix D.2, Tables 14–15, reports the training and cross-task results.

AuthorsOpenReview ↗

Response (Part 2)

C2

(W2) Theorem 2 uses exponential of Lipschitz integral where Lipschitz constant is of learned velocity, and a related constant involving double exponential. [...] Constants for terminal penalty Lipschitz are not quantified for smoothed penalty with delta 0.01. No numerical evaluation of bound is given. Reference assumption that population flow-matching minimizer realizes exact data path requires exact marginal matching by population minimizer, idealized.

(Q2) What are empirical values of Lambda and C constants for final checkpoints? Is exponential on order of tens, thousands, or larger? If large, how to interpret Theorem 2?

Response.

We appreciate the reviewer’s question about the interpretation of Theorem 2. Its purpose is to identify how velocity mismatch and map-consistency error affect generation quality and cross-schedule agreement, rather than to provide a numerically calibrated bound for a trained checkpoint. This analysis motivates retaining both FM and LSD in the training objective, while the stability constants determine the amplification factors in the resulting error upper bounds.

  • Stability factors. Here ℓθ(t)\ell_\theta(t) is the spatial Lipschitz constant of the learned velocity at time tt.

    The stability constants are Cθ=∫01exp⁡(2∫t1ℓθ(u) du) dtC_\theta=\int_0^1\exp(2\int_t^1\ell_\theta(u)\,du)\,dt and Λθ=∫01ℓθ(t)dt\Lambda_\theta=\int_0^1\ell_\theta(t)dt, so 1≤Cθ≤eΛθ1\le\sqrt{C_\theta}\le e^{\Lambda_\theta}. Although the definition of CθC_\theta contains an exponential with a factor of two, taking Cθ\sqrt{C_\theta} brings its contribution to the same exponential scale as eΛθe^{\Lambda_\theta}. For fixed velocity mismatch and map residual, larger stability factors increase the error upper bound; they do not establish that the actual output error is larger. The bound therefore motivates using FM and LSD to reduce the two error terms that contribute to this upper bound.

  • Terminal-loss constant. Smoothing does not increase the Lipschitz constant. Let q(x)≥0q(x)\ge0 measure constraint violation and suppose it is LL-Lipschitz. Our smoothed loss is ρδ(x)=q(x)2+δ2−δ\rho_\delta(x)=\sqrt{q(x)^2+\delta^2}-\delta. Since the derivative of this transform with respect to qq is q/q2+δ2≤1q/\sqrt{q^2+\delta^2}\le1, we have Gρ≤LG_\rho\le L, independently of δ\delta, including δ=0.01\delta=0.01. On Six Disks, qq is the distance to the feasible set, so Gρ≤1G_\rho\le1. If the loss is evaluated through an LDL_D-Lipschitz decoder, the bound becomes LLDLL_D.

  • Population reference assumption. The exact marginal-path assumption specifies the reference population FM flow used to bound the discrepancy between the generated and target distributions. It allows Eq. (13) to relate velocity mismatch to excess population FM risk. This assumption concerns the reference flow, not exact matching of the data path by the trained model. The cross-schedule bound in Eq. (14) compares different compositions of the same learned model and does not require this assumption.

Revised Sec. 5 clarifies the theorem’s training implications and population reference assumption. Appendix B.3.2 gives the smoothing and decoder bounds above.

AuthorsOpenReview ↗

Response (Part 3)

C3

(W3) Direct map baseline retains flow matching and consistency but backprops terminal loss only through one-step map. Because map and velocity share network, direct loss still changes velocity. Comparison isolates added explicit adjoint update, not direct vs adjoint alone. Lambda sweep only on Six Disks validation, three seeds, no search for violation and consistency weights. Removing consistency lowers violation at small K but raises SWD [...] suggesting consistency is fidelity regularizer, not violation penalty, but trade-off not disentangled.

Response.

We agree that the comparison measures the benefit of adding explicit map-adjoint feedback to Direct map. We have clarified this interpretation and added separate direct-penalty and LSD weight ablations. Together with the submitted ablation, these results show how LSD helps preserve fidelity during constraint alignment.

  • What the comparison isolates. Direct map already updates the shared velocity network through the endpoint loss. Both methods retain the same FM, LSD, and direct terminal-loss terms; Ours adds the explicit map-adjoint update. Thus the comparison isolates this additional feedback, rather than all direct versus adjoint effects on the velocity.

  • Weight settings and separate ablations. Direct map and Ours use the same direct-penalty and consistency weights, both fixed at 1. The submitted three-seed Six Disks sweep varies λadj\lambda_{\rm adj} and removes LSD at one fixed adjoint weight. To examine the two other coefficients separately, we trained four Six Disks variants from the same FMM parent for 10,000 updates, changing only the effective direct coefficient or only the LSD coefficient by a factor of 0.30.3 or 33. The effective adjoint and FM coefficients, optimizer, and sampling settings remain fixed. These new ablations use one training seed and 4,096 paired validation samples per KK. The table gives equal-weight means over K=1,2,4,8,16,32,64,128K=1,2,4,8,16,32,64,128.

    Changed coefficientMean violation (×10−3\times10^{-3})SWD
    Reference setting0.6860.0877
    Direct ×0.3\times0.30.6480.1001
    Direct ×3\times30.6780.0640
    LSD ×0.3\times0.30.5460.0850
    LSD ×3\times30.9800.0589

    Increasing the direct coefficient improves average SWD, but its violation effect depends on KK: it decreases violation at K=1K=1 and increases it at K=4K=4. Increasing LSD to 3×3\times improves average SWD while increasing violation; reducing it to 0.3×0.3\times improves both averages in this run. These comparisons separate the effects of the two coefficients. The submitted models retain their original weight settings.

  • What this shows about LSD. LSD penalizes map–velocity inconsistency, not constraint violation. Its fidelity role is particularly clear when it is removed: at λadj=0.3\lambda_{\rm adj}=0.3, SWD rises from 0.0600.060 to 0.1700.170 at K=1K=1 and from 0.0400.040 to 0.0960.096 at K=4K=4, even though violation can improve at small KK (submitted Fig. 4(b)). Together with the separate weight ablations above, this supports retaining LSD to preserve generation quality while adjusting the balance between fidelity and violation.

Revised Sec. 6.3 clarifies the shared-network comparison. Appendix D.6, Table 27, reports the separate weight ablations and their effects on fidelity and violation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary: Contributions and Responses to Concerns

We thank the Area Chair and reviewers for their careful evaluation. We summarize the contribution and the evidence added during the discussion.

  • Simulation-free alignment for any-step constrained generation. Our work addresses the cost of obtaining terminal feedback through ODE rollouts and the difficulty of retaining endpoint alignment when shorter maps are composed (highlighted by Jomt and qHHt). We use a learned-map adjoint to provide efficient feedback and jointly train the local dynamics and finite-time maps, so that alignment is learned before sampling begins. The reviews recognize the efficiency of this adjoint construction (Jomt, fK3u) and the importance of combining training-time constraint alignment, any-step generation and rollout-free training (fK3u).

  • Evidence addressing concerns on feedback reliability and generality. We added parameter-gradient comparisons during alignment and across tasks, directly evaluating the training-feedback direction. At the aligned cross-task checkpoints, cosine is positive against every tested Euler resolution. Additional schedule tests show that the Six Disks violation advantage persists under nonuniform sampling and through K=128K=128. Larger POD representations and paired reconstruction measurements further examine the physical-task results and distinguish reconstruction effects from generated-field residuals. Physical-space correction also demonstrates attainment of the evaluated constraint tolerances with a small measured change in distributional quality.

  • A clearer account of the method and its scope. The revision clarifies baseline comparisons, loss-weight selection and what the theoretical bounds establish. It also explains the distinction from value-based methods that require training rollouts and from guidance applied at inference. These changes connect the problem, training design and supporting evidence more directly, with constraint satisfaction and distributional quality evaluated together.

The revised manuscript incorporates these additions, with changes marked in purple. Thank you for considering our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.