acceptodds
Under review as a conference paper at ICLR 2027

Directional Phase Transitions in Reward Guidance for Flow models

Abstract

Recent advances in reward-guided generation with flow models have demonstrated remarkable performance in steering outputs toward desired reward functions. In particular, inference-time alignment methods are drawing attention, as they adjust the sampling trajectory without extensive retraining of the pretrained model. Recent analyses of sampling trajectories, however, reveal that they are not dynamically homogeneous but pass through distinct phases. Yet, the impact of these phases on inference-time reward alignment remains under-explored. To bridge this gap, we introduce the directional expansion rate, a metric that quantifies how the velocity field of a flow model amplifies or attenuates perturbations toward higher reward. We observe a consistent contraction–expansion–contraction pattern across distinct tasks: data assimilation, PDE solving. To explain this behavior, we derive the phase structure in a closed-form expression for a Gaussian mixture model, establishing a formal link to mode commitment. Based on this analysis, we propose an adjoint-based control strategy that operates exclusively within the expansion window to steer mode selection before the commitment occurs. Across the tasks, we demonstrate that the proposed method generates better-aligned samples with the target reward function.

Primary area:
generative models
Keywords:
generative modeling, controllable generation, flow matching

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_B9iwOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies when to apply reward guidance during flow-model sampling. It measures how the model responds to changes along the reward direction, then applies guidance during the period it identifies as most useful. For hard-constraint PDE tasks, it also projects the sample onto the feasible set later in sampling. The experiments cover PDE sampling and data assimilation. They show improvements on several tested benchmarks.

Critical strengths

1、The paper asks a useful question: when should guidance be applied ?

2、The directional measure offers a different view from measuring overall expansion or contraction.

3、The method combines guidance with late-stage projection for hard physical constraints.

4、Results on the tested scientific tasks are promising, including a longer Navier–Stokes rollout

Critical weaknesses

1、 All experiments are on scientific flow models. The paper shows physical fields, but does not test image or video generation models.

2、The theory uses a simple two-mode Gaussian model. The paper does not show that the same “mode commitment” explanation applies to its PDE and data-assimilation tasks.

3、The effect of the guidance window is hard to isolate. The PDE method also changes when projection is applied. Compared with PCFM, this means the results do not show which change drives the gains.

4、The data-assimilation test set is small. Each task uses eight trajectories. The SEVIR gains over FlowDAS are small relative to the reported variation.

5、The method uses adjoint optimization and repeatedly estimates its guidance window. The paper does not report the resulting runtime or compute cost.

Questions for authors
  1. Is there an existing comparison that changes only the guidance window while keeping the rest of the method fixed?

  2. Can the authors report the runtime or computational cost of the proposed method and main baselines?

  3. On how many of the eight test trajectories does the proposed method outperform FlowDAS?

2 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer B9iw [1/2]

Q1.

Thanks for the thoughtful comment. To isolate the contribution of the window, we present the result for an ablation study.

[Data Assimilation] We first change the window’s boundaries, and move the window to the start point(t=0), the end point(t=1), or a random position. "Controlled steps" counts the guided steps (out of 20) per assimilation step, averaged over the 10 rollout steps. We report RMSE for first k∈{3,6,10}k \in \{3,6,10\} steps, to present errors over longer horizons.

ProposedFull [0,1)[0,1)[te,tc)[t_e, t_c)[tpeak,1)[t_{\rm peak}, 1)Random WindowFixed Window, start at t=0t=0Fixed Window, end at t=1t=1
NSE (SR 8×)
Controlled steps / frame4.4203.413.44.44.44.4
RMSE kk=30.2500.3550.2390.3560.3800.2890.602
RMSE kk=60.2830.4150.4410.4150.4880.4040.718
RMSE kk=100.3050.4250.5610.4260.5340.5090.732
SEVIR (10% sparse)
Controlled steps / frame8.12010.712.88.18.18.1
RMSE kk=30.0640.1710.0710.0680.0940.2200.080
RMSE kk=100.1080.3670.1610.1070.1750.4250.155
[email protected] kk=10 ↑0.5530.2390.4470.5370.4240.2470.392

We summarize the key observations that support our method as follows.

  • Full [0,1)[0,1): Applying more control does not yield better performance. Full-trajectory control uses more controlled steps, yet it performs clearly worse on both tasks.
  • tet_e: Starting the window at tpeakt_{\mathrm{peak}} is critical for long rollouts. While starting at tet_e is competitive at early horizons on NSE, its performance degrades as the rollout length increases.
  • [tpeak,1)[t_{\rm peak}, 1): Closing the window at tct_c is beneficial. Leaving the window open after tct_c degrades performance on NSE. For SEVIR, it merely matches our performance while requiring more controlled steps.
  • Random/Fixed Window: Our method outperforms both cases, verifying that the position of the window matters, not just its length.

[PDE] We conduct the ablation on the Burgers equation benchmark. The comparison has three parts: (i) removing one component at a time (adjoint guidance or projection after tct_c). (ii) changing the window boundaries [te,tc)[t_e, t_c), and (iii) full-trajectory control [0,1)[0,1),

ProposedNo projectionNo adjoint guidance[te,tc)[t_e, t_c)Full [0,1)[0,1)
MMSE0.6360.6759.470.6561.036
SMSE0.6071.1721.510.6181.077
FPD0.961.5422.51.031.31
Mass residual076.130051.78
IC residual0356.360068.32

We summarize key observations as follows.

  • Both projection and adjoint control are necessary for hard constraints. Both projection and adjoint control are necessary for hard constraints. Late projection alone satisfies the constraints but remains far from the target, resulting in high MMSE and FPD. Conversely, removing projection keeps MMSE close to ours, but violates the constraints and degrades SMSE.
  • More control is not better. Full-trajectory control performs worse across all metrics and fails to satisfy the constraints, which is consistent with the DA results.
  • Window start time (tpeakt_{\rm peak} and tet_e). Starting at tet_e is similar with ours on Burgers. and the advantage of tpeakt_{\rm peak} appears mainly in long-horizon DA rollouts.
AuthorsOpenReview ↗

Rebuttal for Reviewer B9iw [2/2]

Q2

We reports the per-sample wall-clock time for all methods. Our method shows a large runtime variance across samples which is caused by the trajectory-adaptive window. The length of the controlled segment, and hence the optimization cost, differs across samples.

Table 1. Per-sample wall-clock time in seconds.

FamilyMethodHeatBurgersRD1DOptimization budget (iterations × steps)
Optimization-basedD-Flow11.79 ± 0.3711.98 ± 0.3912.35 ± 0.371000
OC-Flow14.18 ± 0.3014.70 ± 0.3214.50 ± 0.321000
Projection-basedECI8.41 ± 0.0421.97 ± 0.1626.47 ± 0.06–
PCFM12.07 ± 0.2547.11 ± 0.5826.56 ± 0.36–
BothOurs7.91 ± 1.0255.64 ± 8.7437.95 ± 2.36150 / 600 / 600

We first compare the execution time. Since our method combines adjoint optimization with projection, we compare it against both families of baselines. Our method is the fastest on Heat. On Burgers and RD1D, its runtime is comparable to that of the projection-based baselines: it takes 1.18× and 1.43× the time of PCFM, on which our PDE pipeline is built. It is, however, slower than the optimization-based baselines D-Flow and OC-Flow on these two tasks.

Importantly, the performance gains of our method are not obtained by spending more optimization. When we compare the optimization budget for the adjoint updates (iterations × ODE steps), D-Flow and OC-Flow use 1000 step-updates on every task, whereas ours uses only 150 / 600 / 600 on Heat / Burgers / RD1D. Despite this smaller budget, our method outperforms both optimization-based baselines. This indicates that the gains come from when the adjoint-based control is applied rather than from how much optimization is performed. In addition, our method intervenes on a shorter portion of the trajectory, since the control is applied only inside the window [t̂_peak, t̂_c), while D-Flow and OC-Flow optimize over the entire trajectory.

We will include this table and discussion in the revised manuscript.

Q3

We thank the reviewer for this question. According to the comment, we count the number of improved trajectory by our method compared FlowDAS. As shown in the table, all 8 trajectories on both NSE tasks and SEVIR are improved by our method. This shows that the improvement shared by nearly all trajectories, not from a few outliers.

NSE SR 8×NSE SO 5%SEVIR SO 10%
Improved trajectory8/88/88/8
Reviewer_AuycOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines when to apply inference-time reward guidance in flow models. The authors evaluate the directional expansion rate of the velocity field’s Jacobian along the reward direction computed through the adjoint equation. They identify a contraction, expansion, and recontraction pattern and investigate its connection to mode formation using a symmetric two-component Gaussian mixture model. Their method applies adjoint-based control from the peak directional expansion rate to the subsequent zero crossing and uses late-stage projection for hard-constrained tasks. Experiments show improvements on three PDE benchmarks and Navier–Stokes and SEVIR data assimilation tasks. The appendix provides additional results for multi-step and long-range rollouts. The central contribution is a criterion for intervention timing based on reward-related directional dynamics. However, the available evidence does not clearly establish that this window-selection criterion accounts for the performance gains.

Critical strengths
  • The proposed metric captures reward-related directional dynamics that divergence can obscure. An analytically tractable model connects local expansion to modal structure, offering a useful dynamical interpretation of limited-interval guidance.
  • The method improves prediction accuracy on three PDE systems while satisfying constraints within the reported tolerance. Long-range Navier–Stokes assimilation results also show reduced error accumulation.
Critical weaknesses
  • Comparisons under matched implementations and budgets are missing for full-trajectory control, fixed windows, expansion-onset windows, and the proposed peak-starting window. The contribution of late-stage projection is also not isolated, leaving the source of the gains unclear.
  • Instantaneous directional expansion does not establish terminal reward sensitivity or justify starting control at the peak. Equation (6) is approximate, tct_c is defined implicitly by a trajectory crossing, and nonlinear projection requires conditions to ensure feasibility and mode preservation.
Questions for authors
  • Can the authors compare full-trajectory control, a validation-selected fixed window, [te,tc)[t_e,t_c), and [tpeak,tc)[t_{\rm peak},t_c) using the same optimizer, time grid, and computational budget?
  • What connects instantaneous λw(t)\lambda_w(t) to the effectiveness of an intervention on the terminal output, and what supports starting control at the peak? Can finite-time perturbation responses or reward gains per unit control cost validate the predicted window?
  • How does late-stage projection perform on its own in the PDE tasks? When the projection strategy is held fixed, what additional gains come from the proposed windowed control?
  • Should Equation (6) be identified as an approximation and tct_c as an implicitly defined trajectory crossing time? What parameter and grid conditions justify neglecting terminal re-expansion? If stronger exact results are available, what assumptions do they require?
7 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer Auyc [1/3]

[Q1, Q3]

Per reviewer's comment, we present an ablation study to isolate the contribution of the proposed method.

[Data Assimilation] We first change the window’s boundaries, and move the window to the start point(t=0), the end point(t=1), or a random position. "Controlled steps" counts the guided steps (out of 20) per assimilation step, averaged over the 10 rollout steps. We report RMSE for first k∈{3,6,10}k \in \{3,6,10\} steps, to present errors over longer horizons.

ProposedFull [0,1)[0,1)[te,tc)[t_e, t_c)[tpeak,1)[t_{\rm peak}, 1)Random WindowFixed Window, start at t=0t=0Fixed Window, end at t=1t=1
NSE (SR 8×)
Controlled steps / frame4.4203.413.44.44.44.4
RMSE kk=30.2500.3550.2390.3560.3800.2890.602
RMSE kk=60.2830.4150.4410.4150.4880.4040.718
RMSE kk=100.3050.4250.5610.4260.5340.5090.732
SEVIR (10% sparse)
Controlled steps / frame8.12010.712.88.18.18.1
RMSE kk=30.0640.1710.0710.0680.0940.2200.080
RMSE kk=100.1080.3670.1610.1070.1750.4250.155
[email protected] kk=10 ↑0.5530.2390.4470.5370.4240.2470.392

We summarize the key observations that support our method as follows.

  • Full [0,1)[0,1): Applying more control does not yield better performance. Full-trajectory control uses more controlled steps, yet it performs clearly worse on both tasks.
  • tet_e: Starting the window at tpeakt_{\mathrm{peak}} is critical for long rollouts. While starting at tet_e is competitive at early horizons on NSE, its performance degrades as the rollout length increases.
  • [tpeak,1)[t_{\rm peak}, 1): Closing the window at tct_c is beneficial. Leaving the window open after tct_c degrades performance on NSE. For SEVIR, it merely matches our performance while requiring more controlled steps.
  • Random/Fixed Window: Our method outperforms both cases, verifying that the position of the window matters, not just its length.

[PDE] We conduct the ablation on the Burgers equation benchmark. The comparison has three parts: (i) removing one component at a time (adjoint guidance or projection after tct_c). (ii) changing the window boundaries [te,tc)[t_e, t_c), and (iii) full-trajectory control [0,1)[0,1),

ProposedNo projectionNo adjoint guidance[te,tc)[t_e, t_c)Full [0,1)[0,1)
MMSE0.6360.6759.470.6561.036
SMSE0.6071.1721.510.6181.077
FPD0.961.5422.51.031.31
Mass residual076.130051.78
IC residual0356.360068.32

We summarize key observations as follows.

  • Both projection and adjoint control are necessary for hard constraints. Both projection and adjoint control are necessary for hard constraints. Late projection alone satisfies the constraints but remains far from the target, resulting in high MMSE and FPD. Conversely, removing projection keeps MMSE close to ours, but violates the constraints and degrades SMSE.
  • More control is not better. Full-trajectory control performs worse across all metrics and fails to satisfy the constraints, which is consistent with the DA results.
  • Window start time (tpeakt_{\rm peak} and tet_e). Starting at tet_e is similar with ours on Burgers. and the advantage of tpeakt_{\rm peak} appears mainly in long-horizon DA rollouts.
AuthorsOpenReview ↗

Rebuttal for Reviewer Auyc [2/3]

Q2.

We thank the reviewer for the valuable comment. We first clarify the link between instantaneous rate λw\lambda_w and the terminal reward.

Let gt:=∇xtr(x1)g_t := \nabla_{x_t} r(x_1), where x1x_1 is obtained by solving the sampling ODE x˙t=v(xt,t)\dot x_t = v(x_t,t) from xtx_t, and let wt:=gt/∥gt∥2w_t := g_t/\Vert g_t\Vert_2 and Jt:=∇xv(xt,t)J_t := \nabla_x v(x_t,t). The gradient gtg_t satisfies the adjoint equation g˙t=−Jt⊤gt\dot g_t = -J_t^{\top} g_t. Hence,

ddtlog⁡∥gt∥2=gt⊤g˙t∥gt∥22=−wt⊤Jt wt=−λw(t).\frac{d}{dt}\log\Vert g_t\Vert_2 = \frac{g_t^{\top}\dot g_t}{\Vert g_t\Vert_2^2} = -w_t^{\top} J_t\, w_t = -\lambda_w(t).

Here, ∥gt∥2\Vert g_t\Vert_2 measures the sensitivity of the terminal reward to an intervention at time tt, and λw\lambda_w captures the instantaneous rate at which this log-sensitivity changes with respect to the injection time.

To be specific, as the velocity field expands the space around the trajectory, a perturbation of fixed size becomes relatively smaller; as the field contracts this space, it becomes relatively larger. Hence, the effect of intervention is reduced when λw>0\lambda_w > 0 and increases when λw<0\lambda_w < 0, so the zero-crossings tet_e and tct_c should appear as the local maximum and minimum of Δr(t)\Delta r(t).

In Appendix D.3 in the draft, we added results that empirically support this behavior. On 160 unguided trajectories of the RD1D benchmark, we inject ϵ wt\epsilon\, w_t at t∈{0,0.05,…,0.95}t \in \{0, 0.05, \dots, 0.95\} with ϵ∈{0.01,0.015,0.02}\epsilon \in \{0.01, 0.015, 0.02\}, integrate the unguided ODE to t=1t=1, and measure the relative change Δr/∣r∣\Delta r/|r|. As predicted, for ϵ=0.01\epsilon = 0.01 and 0.0150.015, the response peaks at t=0.10t = 0.10, the first grid point after tet_e. It then decreases through the expansion phase, reaches its minimum just before tct_c, and rises again afterwards. At ϵ=0.02\epsilon = 0.02, the perturbation exceeds the first-order regime and overshoots, so the decrease flattens and the response turns negative near t=0t = 0.

AuthorsOpenReview ↗

Rebuttal for Reviewer Auyc [3/3]

Q4.

We thank the reviewer for the comment.

1. Regarding Eq. (6). Following the comment, we will write Eq. (6) with "≈\approx" and state its regime. The exact onset condition is Eq. (24) in the Appendix B.1, which is a cubic equation in tt. However, its exact root does not admit an interpretable closed form. We therefore simplify it in the regime tσ≪1−tt\sigma \ll 1-t (early stage of small intra-mode variance), which yields the closed form in Eq. (6).

Per the reviewer's comment, we will also clarify the approximation error. We rewrite Eq. (24) and its approximation as gσ(t):=∥μ∥2t(1−t)−(1−t−tσ2) ρt2=0,ρt2=(1−t)2+t2σ2,g_\sigma(t) := \|\mu\|^2 t(1-t) - (1-t-t\sigma^2)\,\rho_t^2 = 0, \qquad \rho_t^2 = (1-t)^2 + t^2\sigma^2, gapp(t):=∥μ∥2t(1−t)−(1−t)3,g_{\rm app}(t) := \|\mu\|^2 t(1-t) - (1-t)^3, where Eq. (6) is the root of gappg_{\rm app}. Then Δg:=gσ(t)−gapp(t)=tσ2[(1−t)(1−2t)+t2σ2].\Delta g := g_\sigma(t) - g_{\rm app}(t) = t\sigma^2\big[(1-t)(1-2t) + t^2\sigma^2\big]. Therefore, the error of the approximation is O(σ2)O(\sigma^2).

2. tct_c is an implicitly defined, trajectory-dependent crossing time. We agree and will state this explicitly. Unlike tet_e and tσt_\sigma, which are functions of (∥μ∥,σ)(\|\mu\|,\sigma) alone, tct_c is the first exit of the trajectory from the expansion region. The boundary B\mathcal B itself is closed-form as in Eq. 7, but locating the crossing requires the trajectory xt(x0)x_t(x_0).

Since v(x,t)v(x,t) in Eq. 16 contains tanh⁡(tμ⊤x/ρt2)\tanh(t\mu^\top x/\rho_t^2), in which tt and xx are coupled nonlinearly, a closed form solution of the ODE x˙t=v(xt,t)\dot x_t = v(x_t,t) is not available. Accordingly, tct_c has no closed-form expression.

3. Terminal re-expansion. As noted in Section 3.2, the re-expansion interval [tσ,1)[t_\sigma, 1) has length σ2/(1+σ2)\sigma^2/(1+\sigma^2), so it is not resolved when the step size satisfies Δt>σ2/(1+σ2)\Delta t > \sigma^2/(1+\sigma^2). In addition, tσt_\sigma is obtained from the isotropic term of Eq. 5 alone, which assumes that the trajectory has already committed to a mode so that the mode-selection (second) term is negligible. We will clarify this point in the revised manuscript.

Reviewer_UFAVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies when inference-time reward guidance should intervene in flow-model sampling. It introduces the directional expansion rate λw = w⊤Jw, the local stretching rate of the pretrained velocity field along the reward direction obtained by transporting the terminal reward gradient backward with an adjoint ODE, and reports a consistent contraction–expansion–contraction pattern of λw along sampling trajectories. For a two-component Gaussian mixture with a reward that is monotone along the mode axis, the authors derive λw in closed form and obtain expressions for the transition times associated with mode commitment. Building on this, they propose a phase-aware steering window: adjoint-based control is applied only on [tpeak, tc), from the peak of λw until it first becomes non-positive, with an additional Gauss–Newton projection during the final contraction phase for tasks with hard constraints. The method is evaluated on hard-constrained PDE sampling (heat, Burgers, reaction–diffusion) and on data assimilation (Navier–Stokes and SEVIR), where it improves prediction accuracy over OC-Flow, D-Flow, FlowDAS, ECI, and PCFM while exactly satisfying the PDE constraints.

Critical strengths
  1. A clean diagnostic quantity with a well-matched theoretical analysis. λw is a simple, interpretable way to ask the right question about guidance timing — not whether the flow locally expands in aggregate (as divergence-based analyses measure), but whether it amplifies or damps perturbations along the specific direction in which control is applied. The explanation of why divergence remains negative and misses the phase structure (Figure 2b) is convincing. The Gaussian-mixture analysis is carried through completely: the Jacobian decomposition into an isotropic term plus a rank-one mode-selection term, the resulting closed form for λw, and the derived transition times te, tc, tσ with a toy experiment whose estimates closely match the predictions.

  2. Consistent improvements on both evaluated tasks, including exact constraint satisfaction. On the PDE benchmarks the method attains the lowest prediction error on all three systems while matching PCFM's exact satisfaction of both the initial-condition and conservation constraints — an informative contrast with PCFM's project-at-every-step strategy. On data assimilation it achieves the lowest RMSE on all three settings, with margins that grow over autoregressive rollouts (Tables 4–5), which is the regime where intervention timing should matter most.

Critical weaknesses
  1. The central causal claim — that the λw-derived window is responsible for the gains — is never isolated. There is no ablation running the same adjoint controller with different windows: the full interval [0, 1], the proposed [tpeak, tc), [te, tc), or a fixed/random window of matched length. The proposed method also differs from OC-Flow in several other respects at once (restarting from the unguided state xpeak, coarse control grid, per-task optimization settings in Table 3) and from PCFM in both steering and projection timing, so the gains cannot be attributed to the phase-aware window specifically. The more basic validation is likewise missing: λw is an instantaneous local stretching rate, whereas the claim concerns the integrated effect of a perturbation on the terminal state; directly injecting equal-magnitude perturbations at different times and measuring the terminal reward change would test exactly this link, and only a toy illustration (Figure 1c) is offered instead.

  2. The empirical scope is much narrower than the claims. The motivation is framed around reward alignment for flow models in general, but all evidence comes from two scientific applications with analytic, physics-defined rewards (constraint residuals, observation mismatch) and pretrained models taken from prior work. No experiment touches learned reward models, text conditioning, or image generation, where the reward direction is state-dependent and need not stay aligned with a single mode axis as the mixture analysis assumes (constant reward direction, two equal-weighted isotropic modes). The phase-structure evidence is also thin: Figure 2 shows aggregate λw curves for two models without reporting the number of trajectories or their variability, and robustness to different observation operators or reward strengths is untested. In addition, the window starts at tpeak rather than the derived te with no sensitivity analysis, tc is estimated by finite differences (step 0.05) on a coarse grid and refreshed every 10 iterations with no study of estimation noise, several margins are within noise (e.g., SEVIR RMSE 0.060±0.010 vs. 0.063±0.012, CSI 0.823 vs. 0.819, over 8 test trajectories), and computational cost is never compared.

Questions for authors
  1. Could you provide the window ablation under a fixed controller and fixed optimization settings: full-trajectory control on [0, 1], the proposed [tpeak, tc), [te, tc), and a fixed or random window of the same total length, on at least one PDE system and one data-assimilation task? This is the experiment that would isolate the contribution of the λw-derived window.

  2. Can you directly validate that λw predicts intervention effectiveness — e.g., by injecting equal-norm reward-direction perturbations at a range of times on unguided trajectories and plotting the induced terminal reward change against the measured λw(t) profile?

  3. How variable is the contraction–expansion–contraction pattern across trajectories, observation operators, and reward strengths? How many trajectories underlie Figure 2, and what fraction of individual trajectories exhibit both sign changes?

  4. How sensitive are the results to the window-estimation procedure: the finite-difference step (0.05), the coarse time grid, the update interval nc = 10, and the choice of tpeak rather than te as the window's left endpoint?

4 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer UFAV [1/4]

Q1.

Thanks for the thoughtful comment. To isolate the contribution of the window, we present the result for an ablation study.

[Data Assimilation] We first change the window’s boundaries, and move the window to the start point(t=0), the end point(t=1), or a random position. "Controlled steps" counts the guided steps (out of 20) per assimilation step, averaged over the 10 rollout steps. We report RMSE for first k∈{3,6,10}k \in \{3,6,10\} steps, to present errors over longer horizons.

ProposedFull [0,1)[0,1)[te,tc)[t_e, t_c)[tpeak,1)[t_{\rm peak}, 1)Random WindowFixed Window, start at t=0t=0Fixed Window, end at t=1t=1
NSE (SR 8×)
Controlled steps / frame4.4203.413.44.44.44.4
RMSE kk=30.2500.3550.2390.3560.3800.2890.602
RMSE kk=60.2830.4150.4410.4150.4880.4040.718
RMSE kk=100.3050.4250.5610.4260.5340.5090.732
SEVIR (10% sparse)
Controlled steps / frame8.12010.712.88.18.18.1
RMSE kk=30.0640.1710.0710.0680.0940.2200.080
RMSE kk=100.1080.3670.1610.1070.1750.4250.155
[email protected] kk=10 ↑0.5530.2390.4470.5370.4240.2470.392

We summarize the key observations that support our method as follows.

  • Full [0,1)[0,1): Applying more control does not yield better performance. Full-trajectory control uses more controlled steps, yet it performs clearly worse on both tasks.
  • tet_e: Starting the window at tpeakt_{\mathrm{peak}} is critical for long rollouts. While starting at tet_e is competitive at early horizons on NSE, its performance degrades as the rollout length increases.
  • [tpeak,1)[t_{\rm peak}, 1): Closing the window at tct_c is beneficial. Leaving the window open after tct_c degrades performance on NSE. For SEVIR, it merely matches our performance while requiring more controlled steps.
  • Random/Fixed Window: Our method outperforms both cases, verifying that the position of the window matters, not just its length.

[PDE] We conduct the ablation on the Burgers equation benchmark. The comparison has three parts: (i) removing one component at a time (adjoint guidance or projection after tct_c). (ii) changing the window boundaries [te,tc)[t_e, t_c), and (iii) full-trajectory control [0,1)[0,1),

ProposedNo projectionNo adjoint guidance[te,tc)[t_e, t_c)Full [0,1)[0,1)
MMSE0.6360.6759.470.6561.036
SMSE0.6071.1721.510.6181.077
FPD0.961.5422.51.031.31
Mass residual076.130051.78
IC residual0356.360068.32

We summarize key observations as follows.

  • Both projection and adjoint control are necessary for hard constraints. Both projection and adjoint control are necessary for hard constraints. Late projection alone satisfies the constraints but remains far from the target, resulting in high MMSE and FPD. Conversely, removing projection keeps MMSE close to ours, but violates the constraints and degrades SMSE.
  • More control is not better. Full-trajectory control performs worse across all metrics and fails to satisfy the constraints, which is consistent with the DA results.
  • Window start time (tpeakt_{\rm peak} and tet_e). Starting at tet_e is similar with ours on Burgers. and the advantage of tpeakt_{\rm peak} appears mainly in long-horizon DA rollouts.
AuthorsOpenReview ↗

Rebuttal for Reviewer UFAV [2/4]

Q2.

We thank the reviewer for the valuable comment. We first clarify the link between instantaneous rate λw\lambda_w and the terminal reward.

Let gt:=∇xtr(x1)g_t := \nabla_{x_t} r(x_1), where x1x_1 is obtained by solving the sampling ODE x˙t=v(xt,t)\dot x_t = v(x_t,t) from xtx_t, and let wt:=gt/∥gt∥2w_t := g_t/\Vert g_t\Vert_2 and Jt:=∇xv(xt,t)J_t := \nabla_x v(x_t,t). The gradient gtg_t satisfies the adjoint equation g˙t=−Jt⊤gt\dot g_t = -J_t^{\top} g_t. Hence,

ddtlog⁡∥gt∥2=gt⊤g˙t∥gt∥22=−wt⊤Jt wt=−λw(t).\frac{d}{dt}\log\Vert g_t\Vert_2 = \frac{g_t^{\top}\dot g_t}{\Vert g_t\Vert_2^2} = -w_t^{\top} J_t\, w_t = -\lambda_w(t).

Here, ∥gt∥2\Vert g_t\Vert_2 measures the sensitivity of the terminal reward to an intervention at time tt, and λw\lambda_w captures the instantaneous rate at which this log-sensitivity changes with respect to the injection time.

To be specific, as the velocity field expands the space around the trajectory, a perturbation of fixed size becomes relatively smaller; as the field contracts this space, it becomes relatively larger. Hence, the effect of intervention is reduced when λw>0\lambda_w > 0 and increases when λw<0\lambda_w < 0, so the zero-crossings tet_e and tct_c should appear as the local maximum and minimum of Δr(t)\Delta r(t).

In Appendix D.3 in the draft, we added results that empirically support this behavior. On 160 unguided trajectories of the RD1D benchmark, we inject ϵ wt\epsilon\, w_t at t∈{0,0.05,…,0.95}t \in \{0, 0.05, \dots, 0.95\} with ϵ∈{0.01,0.015,0.02}\epsilon \in \{0.01, 0.015, 0.02\}, integrate the unguided ODE to t=1t=1, and measure the relative change Δr/∣r∣\Delta r/|r|. As predicted, for ϵ=0.01\epsilon = 0.01 and 0.0150.015, the response peaks at t=0.10t = 0.10, the first grid point after tet_e. It then decreases through the expansion phase, reaches its minimum just before tct_c, and rises again afterwards. At ϵ=0.02\epsilon = 0.02, the perturbation exceeds the first-order regime and overshoots, so the decrease flattens and the response turns negative near t=0t = 0.

AuthorsOpenReview ↗

Rebuttal for Reviewer UFAV [3/4]

Q3.

1. Trajectory-Wise Analysis

For the NSE, we evaluated two observation operators: sparse observations (5% of grid points) and super-resolution (×8\times8). Furthermore, for the PDEs, we conducted an initial condition (IC)-specific analysis, focused on the behavior at t=0t=0.

Regarding figure 2: Following the comment, we have revised it to show the mean ± standard deviation and clarify the settings. In the revised figure, we used reaction–diffusion 1D (RD1D) benchmark, and measured the values over 160 trajectories (16 conditions × 10 random seeds). For data assimilation, we used Navier--Stokes super-resolution ×8\times8 benchmark, and measured over 64 trajectories (8 trajectories 8 random seeds).

Patterns in individual trajectory: (DA) We observed that every trajectory (100%) for both sparse observation 5% (SO5) and super-resolution ×\times 8 (SR8) shows the expansion–contraction pattern. (PDE) We also report the pattern for each constraint(16 constraints × 10 seeds) . As observed in Figure 7 in Appendix D.1, every trajectory (100%) exhibits the expansion–contraction pattern. The initial contraction phase, in contrast, is observed in 78% of trajectories (125/160). This is explained by the behavior at t = 0.

For the exact velocity field at t=0t = 0,

v(x,0)=E[x1]−x,J0=∇xv(x,0)=−I,v(x, 0) = \mathbb{E}[x_1] - x, \qquad J_0 = \nabla_x v(x, 0) = -I,

so λw(0)=−1\lambda_w(0) = -1 for every unit direction ww.

Empirically, however, λw(0)\lambda_w(0) deviates from −1 for some cases. As shown in Figure 8 in Appendix D.1, Since any deviation from −1 at t = 0 is inadmissible for the exact field, these values reflect the approximation error of the learned velocity field v_θ near the endpoint t = 0, rather than a change in the phase structure. By using finite-difference estimation for λw\lambda_w, we suppressed this small-scale deviation by averaged Jacobian, which recovers the analytic value, -1. To be specific, finite-difference estimation of lambda_w becomes:

= \frac{1}{2h}\Big\langle \mathbf w,\ \mathbf v _{\theta} (\mathbf x+h\mathbf w,t)-\mathbf v _\theta(\mathbf x-h\mathbf w,t)\Big\rangle = \frac{1}{2h}\Big\langle \mathbf w,\ \int \mathbf J _\theta(\mathbf x+\tau\mathbf w,t) \mathbf w d\tau\Big\rangle .$$ which is based on the averaged Jacobian. This parallels the finite-size Lyapunov exponent[1], which suppresses the effect of small–scale dynamics. We also have added the analysis on finite-difference estimation in **Appendix D.2**. ### **2. Variability across reward strengths.** For a fixed state along a trajectory, $\lambda_w$ does not depend on the reward strength. It is a quadratic form in the normalized direction $w = \nabla r / \Vert \nabla r \Vert_2$, so rescaling the reward $r$ as $cr$ with $c \neq 0$ leaves $\lambda_w$ unchanged. More generally, the same holds for any reparameterization $\varphi(r)$ with $\varphi' \neq 0$. Since $\nabla(\varphi \circ r) = \varphi'(r) \nabla r$, the direction $w$ changes at most in sign, and the quadratic form $w^\top J w$ is invariant to the sign of $w$. The phase structure is therefore independent of the reward scale. --- ### **Reference** [1] Erik Aurell, Guido Boffetta, Andrea Crisanti, Giovanni Paladin, and Angelo Vulpiani. Predictability in the large: an extension of the concept of Lyapunov exponent. Journal of Physics A: Mathematical and General, 30(1):1–26, 1997. doi: 10.1088/0305-4470/30/1/003.
AuthorsOpenReview ↗

Rebuttal for Reviewer UFAV [4/4]

Q4

Per reviewer's comment, we present a sensitivity study on RD1D benchmark that changes one setting of the at a time. Starting from the reference setting (h=0.05h = 0.05, K=30K = 30, nc=10n_c = 10), we vary the finite-difference step (h∈[0.01,0.1]h \in [0.01, 0.1]), time grid (K∈{15}K \in \{15\}), and the update interval (nc=5n_c = 5, or no update. Control steps refers to the number of time-grid steps inside the window at which the adjoint control is applied.

Proposed (h=0.05h = 0.05)h=0.01h = 0.01h=0.03h = 0.03h=0.07h = 0.07h=0.1h = 0.1K=15K = 15nc=5n_c = 5No update
t^peak\hat t_{\rm peak}0.0790.0540.0700.0920.1050.1010.0800.080
t^c\hat t_c0.4260.4180.4200.4200.4190.4330.4230.513
Control steps10.4110.9310.529.839.414.9810.3113.00
MMSE ↓0.4050.4400.4440.4490.4350.4480.4251.734
SMSE ↓0.2890.3130.3280.3480.3400.2970.35820.947
FPD ↓12.2212.6012.7413.1112.3419.7712.2912.28

We summarize the result as follows:

  • The time grid sets the controller resolution. The window barely moves across grids, but KK changes the number of control steps. A coarser grid (K=15K = 15) halves the control steps inside the window and degrades MMSE and FPD.
  • Ablation on hh. tct_c is insensitive to hh over h∈[0.01,0.1]h \in [0.01, 0.1], while tpeakt_{\rm peak} moves later as hh grows. The metrics change only mildly across hh.
  • Ablation on ncn_c. Updating t^c\hat t_c more often (nc=5n_c = 5) gives similar MMSE and FPD, while SMSE increases. Without updates, tct_c fails to track the transition time, and MMSE and SMSE degrade sharply, as tct_c depends on the trajectory.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.