acceptodds
Under review as a conference paper at ICLR 2027

Classifier-Free Guidance from DDPM to DDIM: Inverse Optimal Control and Terminal Laws

Abstract

Classifier-free guidance (CFG) applies the same score extrapolation in stochastic and deterministic diffusion samplers, yet generally produces different output distributions. We analyze this sampler dependence through inverse optimal control and density transport in the continuous-time, exact-score setting. We identify explicit running terms that make prescribed CFG policies optimal under specified quadratic effort metrics. The resulting terminal-density representations retain initialization and trajectory-dependent corrections to the endpoint power tilt. They connect stochastic reference-path averaging to deterministic guided transport through a zero-noise limit under stated regularity conditions. The characterization extends to smooth time–evidence potentials and motivates coefficient-budget coordinates for schedule selection and Online-, which adapts guidance using a numerical evidence estimate. Analytic Gaussian-mixture experiments support the density identities and demonstrate that matching coefficient budgets does not determine the terminal distribution.On CIFAR-10, budget-based selection improves power-cosine schedules (PCS) within the tested search grids. Online-K reduces mean FID relative to PCS by in paired DDIM-200 evaluations on CIFAR-10, with improvements across all three sampling seeds, and by with pretrained DiT-XL/2 and DDIM-50 on ImageNet 256 × 256.

Primary area:
generative models
Keywords:
Classifier-free Guidance, Optimal Control, Diffusion Model

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_8jvhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies classifier-free guidance (CFG) in stochastic and deterministic diffusion samplers from the perspectives of inverse optimal control and probability density evolution. The authors derive control objectives under which prescribed CFG policies are optimal. They also characterize the terminal distributions of guided sampling and explain how initialization and sampling trajectories affect the final distribution. The theoretical analysis connects stochastic and deterministic sampling through a zero-noise limit. Based on these results, the paper introduces coefficient-budget coordinates for guidance schedule selection and Online-K for adaptive guidance. Experiments on Gaussian mixtures, CIFAR-10, and ImageNet validate the theoretical findings and demonstrate improvements in FID.

Critical strengths
  • A systematic theoretical analysis of CFG. The paper provides a unified perspective on CFG through inverse optimal control and terminal density evolution. The theoretical results clarify how guidance affects stochastic and deterministic sampling, offering useful insights into their different distributional behaviors.

  • Meaningful characterization of terminal distributions. The paper derives terminal density representations that account for initialization and accumulated guidance effects. These results help explain why the final sampling distribution can differ from the density suggested by local score extrapolation. The connection between stochastic and deterministic sampling further strengthens the theoretical contribution.

  • Practical extensions supported by experiments. The paper connects its theoretical analysis to guidance schedule selection and adaptive guidance through Online-K. The experiments include analytical Gaussian mixtures and image generation benchmarks. The reported results show that the proposed approaches can improve FID over the evaluated baselines.

Critical weaknesses
  • Limited novelty and practical significance of the inverse optimal control formulation. The paper constructs control objectives under which prescribed CFG policies become optimal. However, this approach appears closely related to standard inverse optimal control techniques, where the objective is designed to match a given policy. It remains unclear what new insights this formulation provides beyond existing theoretical analyses of CFG. In particular, the derived optimality does not directly establish that CFG is optimal for generation quality or provide a principled way to select guidance schedules. This limits the significance of the theoretical contribution.

  • Insufficient empirical validation of the proposed adaptive guidance mechanism. The theoretical analysis assumes exact scores and continuous-time dynamics, while Online-K relies on estimated conditional evidence and finite-step sampling. More importantly, the experiments do not clearly separate the effect of evidence-based feedback from changes in guidance strength and timing. Therefore, the reported FID improvements cannot be directly attributed to the proposed feedback mechanism. Additional ablations with matched guidance schedules and computational budgets would help establish its effectiveness.

Questions for authors

How does the proposed framework differ from existing time-varying CFG methods? Several recent studies, including β\beta-CFG (Classifier-free Guidance with Adaptive Scaling) and C$^2$FG (Control Classifier-Free Guidance via Score Discrepancy Analysis, CVPR 2026), have investigated adaptive guidance through different theoretical motivations and scheduling strategies. Could the authors further discuss the conceptual and theoretical differences between these methods and the proposed framework? In particular, it would be helpful to clarify what additional insights or practical advantages the inverse optimal control formulation and Online-K provide beyond existing time-dependent guidance approaches.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 8jvh.

Thank you for your constructive comments. We address your concerns below.

1. Relation to beta-CFG, C2FG and What the inverse-control characterization adds

Guidance signals and methodological scope.

β\beta-CFG combines a beta-shaped temporal envelope with current conditional--unconditional prediction-gap normalization, yielding sample-dependent scales (Malarz et al., arXiv:2502.10574, Eq.17). C$^{2}FGusesscore−discrepancyboundstomotivateasharedexponentialtimeschedule(Gaoetal.,arXiv:2603.08155,Section 3.2).Online−FG uses score-discrepancy bounds to motivate a shared exponential time schedule (Gao et al., arXiv:2603.08155, Section~3.2). Online-K$ uses an accumulated estimate motivated by conditional density-ratio dynamics. They differ in scale-setting information and policy characterization, beyond time variation/sample dependence.

What the control and terminal-law characterization adds.

For a prescribed policy, we give an explicit compatible objective, its optimality gap and a sampler-dependent terminal law, exposing initialization and visited-state effects. Stochastic density corrections average over endpoint-conditioned reference paths; deterministic corrections follow a guided characteristic. Identical schedules can thus yield different terminal laws.

For a smooth time-only schedule, Ft=a(t)KtF_t=a(t)K_t gives Rtra=a˙K+(a2−a)E.R_{\mathrm{tr}}^a=\dot aK+(a^2-a)E. This analyzes boundary/trajectory corrections for exponential or smooth beta-shaped envelopes. Timing and encountered score separation prevent coefficient budgets from determining terminal laws. For evidence response a=ϕKa=\phi_K, the residual also contains ∂Ka E\partial_Ka\,E, missed by substituting a varying scale into the constant-scale expression. These results use the stated exact-score assumptions; they neither claim the induced objective measures generation quality nor imply that β\beta-CFG's full norm-dependent rule is always representable as a(t,Kt)a(t,K_t).

What distinguishes the Online-KK feedback signal.

Under exact scores, Kt(x)=log⁡[pt(x∣c)/pt(x)]K_t(x)=\log[p_t(x\mid c)/p_t(x)] is a state function whose gradient is the conditional--unconditional score difference; its norm does not determine KK. Along a guided ODE, ddtKt(Xt)=−Ht(Xt)+(at−1)Et(Xt).\frac{d}{dt}K_t(X_t)=-H_t(X_t)+(a_t-1)E_t(X_t). Divergence in HtH_t prevents score-gap energy alone from recovering KK. Online-KK approximates this evolution for its sigmoid response: equal gap norms at the same timestep can yield different coefficients when accumulated estimates straddle the threshold, unlike time-only or gap-norm rules. With learned scores, zero initialization and finite steps, K^j\widehat K_j is history dependent.

2. Bridge from exact-score theory to finite-step experiments

Analytic mixtures test terminal laws and omission of their corrections. Learned CIFAR-10 tests show sampler and timing dependence without reconstructing high-dimensional terminal-density corrections: at 996 steps, cross-sampler 103MMD^210^3\widehat{\mathrm{MMD}}^2 is 0.2590.259 for w=1w=1, 0.7060.706 for w=3w=3; all six equal-coefficient pulse comparisons have Holm-adjusted p=0.020p=0.020 (Tables L.1, L.3).

Table N.2 isolates quadrature, initialization and trace errors on a common conditional DDIM trajectory with exact mixture scores. At 256 steps, evidence MAE: 3.009×10−53.009\times10^{-5} with exact initialization/divergence, 0.19700.1970 with zero initialization/one probe, and 0.19130.1913 with zero initialization/exact divergence. These validate ideal quadrature; finer grids alone do not remove initialization bias.

3. Matched mean schedule and computational cost

Table M.1 freezes the timestep-wise mean from the original 150K CIFAR-10 trajectories, matching initial noise, labels, model, grid and FP32 arithmetic. Values: mean ± sample SD over three seeds, 50K images/method/seed.

MethodFID ↓Precision ↑Recall ↑
Online-KK2.561740±0.0140380.686147±0.0015510.578007±0.000244
Frozen mean replay2.630493±0.0176260.691433±0.0016480.574553±0.002111

Online-KK achieved lower FID than replay in all three seeds (2.5617 vs. 2.6305 on average), with higher mean recall and lower mean precision.

The low-cost energy proxy reuses conditional/unconditional predictions, avoiding endpoint-estimator backward computation. At fixed DDIM-200/50 on CIFAR/ImageNet, sampling time is 1.002×1.002\times/1.003×1.003\times fixed CFG, matching steps and approximate cost.

DatasetFixed CFGPCSβ\beta-CFGFBGC$^2$FGOnline-K (low-cost)
CIFAR2.947±.0312.605±.0273.087±.0354.526±.0212.877±.0412.658±.035
ImageNet3.738±.0443.614±.0693.596±.0223.678±.0093.460±.0233.508±.015

Proxy mean FID improves on fixed CFG but trails PCS/C$^{2}FGonCIFAR/ImageNet.Qualityusesarchivedevaluations;timingwasmeasuredseparately.EachoffivetunedImageNetmethodsscreens12configurationsat1Kimageseachbeforethree50Kevaluations;CIFARuses18candidatesfororiginalbaselines,sixforOnline−FG on CIFAR/ImageNet. Quality uses archived evaluations; timing was measured separately. Each of five tuned ImageNet methods screens 12 configurations at 1K images each before three 50K evaluations; CIFAR uses 18 candidates for original baselines, six for Online-K$, twelve for C$^{2}$FG. Records include seed provenance.

In Author Response 3, the performance on the ImageNet dataset does not seem to demonstrate the method's advantage, and the comparisons are not conducted under the standard conditions for SDE-250 inference (FID-50K). Therefore, I prefer to maintain the score.

Reviewer_6xCAOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper explains why classifier-free guidance (CFG) produces different output distributions when the same guidance scale is used with stochastic (DDPM) versus deterministic (DDIM) samplers. First, the authors derive objectives for which CFG is the optimal policy. They then derive CFG’s exact output distribution, which consists of the commonly assumed “power tilt” p1−w(pc)wp^{1−w}(p^c)^w multiplied by two corrections: a boundary factor that accounts for where sampling starts, and a weight based on accumulated score disagreement. The sampler dependence comes from where that weight is computed. DDPM averages it over many random paths, while DDIM evaluates it along a single guided trajectory. The authors extend the framework to guidance scales that vary with time and evidence. From this they propose: (i) new coordinates for searching guidance schedules; (ii) Online-K, which adjusts guidance per sample using a running estimate of how class-like the sample is. Toy Gaussian experiments support the theoretical results, and image experiments show modest FID gains on CIFAR-10 and ImageNet.

Critical strengths
  • The paper is well written, if somewhat dense.
  • The theoretical sections (3 and 4) are clear and carefully stated, and they give a principled explanation of CFG’s sampler dependence. In particular, the terminal-density results show why the common belief that “CFG samples the power tilt” fails. The output is the tilt times an initialization factor and a path-dependent weight, and stochastic and deterministic samplers differ in which trajectories that weight is evaluated on.
  • The Gaussian-mixture experiments are well designed, with ablations showing that each correction term is needed.
Critical weaknesses
  • Link from theory to method is weak:
    • No image experiment tests the theory’s central motivation. All image results use DDIM. The predicted DDPM/DDIM differences are never examined beyond toy settings.
    • The (peak, budget) grid is a reparameterization of the same two-parameter PCS family. The CIFAR gain therefore reflects where the 12 grid points were placed, as the authors partly acknowledge.
  • Modest gains: Apart from the toy Gaussian experiments, the empirical evidence is modest. FID gains of 1–3% with three seeds. I think this partly reflects the paper trying to be both a theory paper and a methods paper. The theoretical characterization is the stronger contribution. The practical methods, budget coordinates and Online-K, are only loosely derived from it and show small gains. The paper might be more compelling if it used its experimental section to test the theory’s predictions on real image models, for example the predicted DDPM/DDIM difference or the size of the boundary factor, rather than positioning the methods as headline results.
Questions for authors
  • All the guarantees assume exact scores. Learned conditional and unconditional networks usually don’t correspond to any consistent pair of densities. Thus, the evidence KK may not even be well defined. Do the authors expect the qualitative picture (power tilt plus corrections, and the SDE/ODE difference) to survive with real networks?
  • The evidence interpretation relies on discrete classes. How do the results and Online-K extend to text conditioning?
  • See weaknesses.
2 comments
AuthorsOpenReview ↗

Response to Reviewer 6xCA

Thank you for suggesting us to test the theory's implications on real-image models. Section 6.2 and Appendix L report sampler and guidance-timing tests on a frozen CIFAR-10 model; Appendix N adds evidence diagnostics.These learned-score tests differ from the exact-score characterization.

1.Real-image tests of the theory's predictions

Section 6.2 and Appendix L compare DDIM and DDPM with one frozen CIFAR model, a shared noise table, and 10K images per configuration.Table L.1 reports cross-sampler 103MMD^210^3\widehat{\mathrm{MMD}}^2:

&249&498&996\text{ steps}\\\\\hline w=1&0.727&0.486&0.259\\\\ w=3&0.724&0.811&0.706 \end{array}$$ All six Holm-adjusted $p$-values are $0.020$. The $w=1$ discrepancy decreases with refinement; the $w=3$ discrepancy persists at similar magnitude. Learned scores and finite-noise initialization remain, so the nonzero $w=1$ difference is not attributed to guidance alone. Moving the same pulse among early, middle and late intervals, with matched endpoints, peak, mean, coefficient histogram, and coefficient budget $5/12$, changes output distributions: all six have Holm $p=0.020$ (Table L.3), directly testing guidance-history dependence. Appendix L retains local response prediction and integrated early-to-late schedule-response tests, with all numerical controls and independent population evaluations. These differentiate the implemented sampler and test timing dependence. ## 2. Role of budget coordinates We agree that peak--budget coordinates reparameterize PCS. Appendix I compares two prespecified 12-candidate searches, selected with 10K development images per candidate and evaluated on three fresh paired 50K-image seeds. Table I.1 reports: |Selection|$(m,s)$|FID $\downarrow$|Accuracy (%)|Precision|Recall| |---|---|---|---|---|---| |Original PCS|$(1.3,4)$|$2.6334\pm0.0096$|98.349|0.6877|0.5760| |Parameter grid|$(1.45,8)$|$2.6232\pm0.0116$|98.931|0.6889|0.5716| |Budget grid|$(1.31425,6.03469)$|$2.5993\pm0.0110$|98.459|0.6856|0.5760| FID is mean $\pm$ sample SD; other metrics are seed means, and selected parameters are rounded. The grids span different domains: this is a finite-grid selection result, not evidence of general search superiority. The theory shows that coefficient budgets alone do not determine the terminal law. ## 3. Learned scores and estimated evidence Learned conditional and unconditional score fields need not correspond to temporally consistent densities, nor their difference to the gradient of a global scalar potential. Section 5.3 distinguishes the ideal density ratio $K_t(x)$ from Online-$K$'s history-dependent signal $\widehat K_j$. For the sigmoid response, Eq. (22) gives $$|a(\widehat K)-a(K)|\leq \frac{\lambda\gamma}{4}|\widehat K-K|.$$ This pointwise input-sensitivity bound gives no trajectory or terminal-distribution guarantees. A constant signal offset can be absorbed into the threshold, but tuning does not restore exact-score assumptions. Appendix N.2 separates numerical errors using analytic evidence. Along a common conditional DDIM trajectory with exact mixture scores, exact initialization and divergence give MAE $3.009\times10^{-5}$ at 256 steps. Zero initialization leaves MAE $0.1913$ with exact divergence, or $0.1970$ with one trace probe (Table N.2). Finer quadrature alone does not eliminate initialization bias. With the specified radial score-gap perturbation, selected comparisons give mean sliced-$W_1$ $\pm$ sample SD over five final seeds (DDIM-256, 10K samples per seed; Tables N.3--N.4): |Method|2D eight-component mixture|1D three-component mixture| |---|---|---| |Fixed CFG|$0.11535\pm0.01771$|$0.06442\pm0.03305$| |Linear|$0.09697\pm0.03414$|$0.05606\pm0.02055$| |Mean replay|$0.12140\pm0.02971$|$0.06658\pm0.02986$| |Online-$K$|$0.09897\pm0.01649$|$0.05834\pm0.03400$| Online-$K$ improves mean sliced-$W_1$ over fixed CFG and mean replay in both mixtures, although the linear schedule has a slightly lower mean. These synthetic results establish neither learned-image calibration nor guarantees under arbitrary score errors. Image experiments support the tested sampler and timing dependence. Quantitative model-error control for the exact terminal-law characterization remains future work (Section 7). ## 4. Text conditioning The density ratio does not require finitely many classes. For fixed $c$, including a text prompt, define $$K_t(x;c)=\log\frac{p_t(x\mid c)}{p_t(x)}.$$ The identities apply when these conditional and unconditional density families satisfy the stated assumptions under the same forward process. No prompt-prior evaluation or prompt classifier is required; the discrete-class posterior interpretation is optional. Online-$K$ could use the text-conditional/unconditional score difference in accumulated update. Our experiments are class-conditional, so text-to-image calibration and effectiveness remain untested. An additive evidence-convention change requires a corresponding sigmoid-threshold shift.

Thank you for your reply. The authors have addressed my questions / weaknesses and I will be maintaining my original score.

Reviewer_4MYWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper explains why classifier-free guidance gives different results with stochastic (DDPM) and deterministic (DDIM) samplers. It shows which objective each sampler's guidance actually optimizes and derives the exact output distribution in each case. From this, it builds a schedule-search heuristic and an adaptive guidance method (Online-K), which modestly improve FID on CIFAR-10 and ImageNet.

Critical strengths

(i) The paper gives closed-form output densities for guided DDPM and DDIM linked by a zero-noise limit.

(ii) On Gaussian mixtures with exact scores, the DDIM formula matches the flow-Jacobian density. The DDPM formula matches at w=1 and w=3, dropping the initialization factor raises the error sharply. An equal-budget early / late guidance experiment shows that timing matters, as the theory claims. The authors also openly report where validation is inconclusive (e.g., w=8).

Critical weaknesses

(i) The theory assumes exact scores, but Online-K relies on a learned evidence estimate that the authors themselves describe as uncalibrated. The method therefore runs outside the guarantees. The gains are small and rest on few seeds. The comparisons are also not fully matched. Online-K was tuned while the baseline to my understanding was not re-tuned. Online-K adds compute, and it lowers other metrics (classifier agreement, precision).

(ii) The inverse optimality result is a general construction that applies to any smooth guidance rule. The authors acknowledge that the resulting objective is not unique and does not reflect sample quality. The main new contribution is the exact characterization of the output distributions.

Questions for authors

(1) I was a bit confused with the ImageNet results. It seems that they are described differently in different parts of the paper. In some places they are described as an average over three seeds, in others they read as a single run with a larger improvement. Which is correct, and were all methods evaluated under the same setup, can you provide a per seed table?

(2) Is the gain from Online-K really due to adapting guidance to each sample? Would a simple fixed schedule that mimics Online-K's average guidance over time perform just as well?

(3) How does Online-K compare against a baseline that received the same amount of tuning and the same compute budget?

(4) The paper shows that guidance is optimal for some objective, but this construction seems to work for almost any guidance rule. What concrete insight or design choice depends on this control view, beyond the formulas for the output distributions?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 4MYW

Thank you for the comments. The revision characterizes prescribed CFG policies and sampler-dependent terminal laws; Online-KK is one motivated implementation. We address reporting, mean replay, tuning, cost, and metrics.

1. Consistent ImageNet reporting and per-seed results

We apologize for inconsistent ImageNet reporting. The three comparisons are:

Scope and implementationOnline-K FIDPCS FIDReduction
Original Table 3: energy proxy, three-seed mean3.50793.61572.98%
Original appendix: energy proxy, seed 13.5151273.6932974.82%
Revised Table 3 / K.1: endpoint, three-seed mean3.31493.61408.28%

The revised endpoint uses per-sample accumulated VJP/trapezoidal updates, not a reevaluation of the original energy proxy. PCS was recomputed from supplied seed records; proxy results are separate in the low-cost study. Corrected endpoint and baseline seeds:

MethodSeed 0Seed 1Seed 2
Fixed CFG3.6976783.7850533.729963
PCS3.5650443.6932973.583599
β\beta-CFG3.5888773.6208623.578965
FBG3.6770553.6884093.669823
C$^2$FG3.4377453.4828643.459829
Online-K3.3164883.3148963.313354

Each FID is a separate 50K evaluation; mean and sample SD are across seeds. Methods share the checkpoint, VAE, DDIM-50, and evaluation size. Seed records give source IDs. Random streams are not all paired, so these compare reported configurations.

2. Frozen mean-guidance schedule

The replay test (Table M.1) freezes Online-KK's stepwise mean guidance over the original 150K CIFAR trajectories for every replay sample. Model, grid, initial noise, labels, and evaluation size are matched.

MethodFID (mean ±\pm sample SD)
Online-KK2.561740±0.0140382.561740\pm0.014038
Frozen mean replay2.630493±0.0176262.630493\pm0.017626

Online-KK lowers mean FID by 2.612.61% and all three seeds‘ FID: this mean curve does not reproduce its performance. The curve is retrospective, without an independent calibration split or separate schedule optimization. Replay removes the assosiation between each sample and its guidance curve, and we find out adapting guidance to each sample actually makes improvements in FID performance.

3. Tuning and computational budget

On ImageNet, Online-KK, PCS, β\beta-CFG, FBG, and C$^{2}FGeachscreen12configurationswith1Kimagespercandidate,selectbydevelopmentFID,andfreezeforthree50Kevaluations.Online−FG each screen 12 configurations with 1K images per candidate, select by development FID, and freeze for three 50K evaluations. Online-K$'s grid is

λ∈{1,2},γ∈{1,4},κ∈{1,2,4},\lambda\in\{1,2\},\quad\gamma\in\{1,4\},\quad\kappa\in\{1,2,4\},

selecting (1,1,4)(1,1,4). On CIFAR, original baseline families screen 18 candidates, Online-KK six, and added C$^{2}$FG twelve, with 10K images per candidate.

To assess feedback at approximately fixed grid and elapsed cost, our low-cost variant reuses conditional/unconditional prediction differences as an accumulated energy proxy, avoiding the endpoint estimator's backward computation. All methods use DDIM-200/50 on CIFAR/ImageNet. On A100 GPUs, measured sampling time is 1.002×1.002\times/1.003×1.003\times fixed CFG, respectively.

DatasetFixed CFGPCSβ\beta-CFGFBGC$^2$FGOnline-K (low-cost)
CIFAR2.947$\pm$.0312.605$\pm$.0273.087$\pm$.0354.526$\pm$.0212.877$\pm$.0412.658$\pm$.035
ImageNet3.738$\pm$.0443.614$\pm$.0693.596$\pm$.0223.678$\pm$.0093.460$\pm$.0233.508$\pm$.015

The proxy improves reported FID over fixed CFG on both datasets at near-baseline measured sampling cost. It ranks second by mean FID, behind PCS on CIFAR and C$^{2}FGonImageNet,preservingtheendpointestimator′srelativeadvantagesatlowercost.Thisquality–costtrade−offdoesnotestablishendpointOnline−FG on ImageNet, preserving the endpoint estimator's relative advantages at lower cost. This quality–cost trade-off does not establish endpoint Online-K$'s superiority at matched total wall-clock budget.

Paired CIFAR PCS gives 2.6114±0.01412.6114\pm0.0141 versus endpoint Online-KK's 2.5617±0.01402.5617\pm0.0140, a 1.901.90% reduction. In the main comparisons, Online-KK has lower precision than PCS and recall ranked second on CIFAR and first on ImageNet.

4. Concrete insights from the control view

Verification is standard; the contribution is the explicit CFG-specific objective and gap under specified dynamics and effort metrics:

R_{\rm D}=-wH+\left(\frac{w^2}{2}-w\right)E.$$ The same scale thus has different compatible SDE and ODE objectives. The gap compares controls under one fixed objective, weighting drift deviations along their visited trajectories. For $F_t=a(t)K_t$, the time-only transport residual contains $\dot aK+(a^2-a)E$: coefficient budget alone cannot determine the terminal law. The real-image pulse test checks this: all six timing comparisons differ after Holm correction despite matched coefficient statistics ($p=0.020$; Table L.3). For evidence feedback, the sigmoid bound $$|a(\widehat K)-a(K)|\leq\frac{\lambda\gamma}{4}|\widehat K-K|$$ shows that greater amplitude or sharpness raises the input-error sensitivity bound at a fixed state, motivating joint assessment of estimator and response parameters.
Reviewer_4MYWOpenReview ↗

Response to rebuttal

I thank the authors for their rebuttal. While I appreciate the additional information provided by the authors, I still view this paper as a work in progress. I am still not convinced that the method improves performance under matched compute. As the authors acknowledge themselves "This quality-cost trade-off does not establish endpoint Online-K's superiority at matched total wall-clock budget.". Therefore I maintain my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.