{"id":"c5af143e-f16f-45af-8b0a-79961d3fc911","arxiv_id":"2607.02137","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ART-RL learns adaptive diffusion sampling timesteps via continuous-time control and Gaussian actor–critic RL, improving and transferring over hand-designed grids at matched budgets.","lead":"This paper learns better timestep grids for diffusion-model sampling by treating the sampling clock speed as a control and training it with continuous-time actor–critic RL. If the method holds up, existing image generators can improve quality at the same compute budget by swapping only the timestep list.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Euler-local |Q|θ² objective is the load-bearing allocation signal, yet main gains and transfer claims are measured under Heun/RK4 without a matching higher-order residual analysis.","rationale":"The reader correctly isolates the Euler surrogate vs Heun/RK4 gap as the weakest assumption behind the strongest claim. The control/RL stack (ART \to Gaussian ART-RL equivalence, policy evaluation/improvement, actor–critic updates) is internally coherent under the stated objective; the soft spot is whether that objective remains the right allocation principle once the sampler is no longer Euler. Experiments already show gains under Euler (Table 3) and under Heun/RK4, and transfer is broad, so the concern does not overturn the paper—it keeps the verdict CONDITIONAL until a solver-matched residual (or a clear ablation showing the Euler proxy is sufficient) is provided. No stronger internal inconsistency appears in the derivations or in the controlled comparisons that hold score model, solver, and pipeline fixed. Reproducibility and tiny high-NFE margins are secondary. Thus I agree with the reader’s load-bearing concern and leave the verdict unchanged.","tokens_in":25351,"tokens_out":868,"duration_ms":8537,"concrete_test":"On CIFAR-10 EDM with fixed pretrained score and Heun sampler, retrain ART-RL under two objectives: (i) the paper’s Euler |Q|θ^{2} cost, and (ii) an analogous one-step residual cost obtained by expanding the Heun map to the same order (or a finite-difference estimate of local Heun truncation along trajectories). Distill both to time-only grids at K∈{5,10,18}, evaluate FID at matched NFE=2K-1 over three seeds of 50k samples, and transfer both grids to AFHQv2/FFHQ at the same NFEs. If the Heun-aligned objective does not improve (or reverse) the FID margins vs EDM/DPM, and transfer gaps stay within ~0.05 FID of the Euler-trained grids, the surrogate–solver mismatch is not load-bearing; otherwise the central drop-in/transfer claim needs a solver-matched objective.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim is that ART-RL grids, learned from the continuous-time control objective and plugged into existing samplers by changing only the timestep grid, consistently beat Uniform/DPM/EDM at matched budgets and transfer without retraining. That claim rests on the objective in §3.2: the leading-order one-step Euler residual E_i = (h_i^{2}/2)θ_i^{2} Q + O(h_i^{3}) (Eq. 7), with Q given by (8), which is turned into the running cost |Q|θ^{2} in J_\theta (Eq. 9). Theorems 1–3 and Algorithm 1 then optimize that cost under Gaussian policies. The main image results, however, evaluate Heun (CIFAR-10 EDM, NFE=2K-1; Tables 2,5–7) and RK4 (MNIST, NFE=4K-3; Table 4), and the transfer suite reuses the same CIFAR-10-trained schedule under those higher-order solvers and under EDM2 latent sampling. The paper never derives an analogous local residual for Heun/RK4, nor shows that the Euler stiffness field Q remains the correct allocation signal once the integrator order changes. Section 7 lists this as a limitation, but the headline transfer/generalization claims treat the learned grids as solver-agnostic. If the Euler surrogate systematically misallocates steps relative to the true local truncation of the deployed solver, the reported FID/LeNet-FID gains and cross-budget/dataset/pipeline transfer would be empirical coincidences rather than consequences of the control theory. Distillation to time-only schedules further removes any residual state dependence that might have compensated for the mismatch.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper formulates timestep allocation for score-based diffusion sampling as Adaptive Reparameterized Time (ART): a continuous-time control problem in which the control is the speed of a reparameterized sampling clock, so that a uniform grid on the learned clock induces nonuniform physical-time steps under a fixed budget. The objective is a leading-order Euler residual surrogate |Q|θ² plus a Lagrange term for the time budget (Eqs. 7–9). To solve the high-dimensional deterministic problem, the authors introduce ART-RL with Gaussian policies whose variance is proportional to 1/|Q|, prove that the mean of the optimal Gaussian policy recovers the ART optimizer (Theorem 1), and develop continuous-time policy evaluation/improvement and martingale orthogonality conditions that yield implementable actor–critic updates (Theorems 2–3, Algorithm 1). Empirically, distilled ART-RL grids are plugged into existing samplers by changing only the timestep schedule and improve FID/LeNet-FID over Uniform, DPM, and EDM at matched NFE on a 1D analytical-score task, MNIST, CIFAR-10 (Euler and Heun), and transfer without retraining across budgets, AFHQv2/FFHQ/ImageNet-64, and EDM2 ImageNet-512.","tokens_in":25783,"tokens_out":1522,"duration_ms":14443,"significance":"If the results hold, this is a genuine contribution: a control-theoretic alternative to hand-crafted diffusion timestep grids, with a clean optimizer-level equivalence between deterministic ART and Gaussian ART-RL, and with actor–critic theory specialized to time reparameterization. The experimental design is strong for the claim that only the grid changes—score model, backbone, solver, and pipeline are held fixed—and the transfer suite (cross-budget, cross-dataset, EDM→EDM2 latent, high resolution) is unusually thorough. Distillation to a fixed precomputed grid makes the method practically drop-in with no inference overhead. The main open scientific question is whether the Euler-local surrogate remains the right allocation signal under the higher-order solvers used in the headline image results; that is a limitation the paper itself flags, not a circularity or construction artifact.","major_comments":[{"comment":"§3.2 Eqs. (7)–(9) and §6.3–6.5: The control objective is the leading-order one-step Euler residual E_i = (h_i²/2) θ_i² Q + O(h_i³), with running cost |Q|θ². The main image and transfer claims, however, evaluate Heun (CIFAR-10 EDM, NFE=2K−1; Tables 2, 5–7) and RK4 (MNIST, NFE=4K−3; Table 4), and reuse the same CIFAR-10-trained schedule under those solvers and under EDM2. There is no matching local residual for Heun/RK4, nor a controlled ablation that isolates whether Q remains the correct stiffness field once the integrator order changes. §7 lists this as a limitation, but the abstract and transfer claims treat the grids as solver-agnostic. Please either (i) derive or empirically validate a higher-order residual and retrain under it for the Heun/RK4 settings, or (ii) substantially qualify the solver-transfer claims and report a same-solver vs cross-solver comparison that shows how much of","section":null},{"comment":"§6.2–6.3 and distillation: After training, policies are distilled to time-only mean θ curves (Figs. 4–5, Tables 1–3). The theory (Theorems 1–3, dynamics (4)) allows state-dependent feedback θ(t,x,ψ). The paper reports that confidence bands are narrow, but does not quantify the performance gap between the full feedback policy and the distilled time-only grid on CIFAR-10/ImageNet, nor when residual state dependence would matter (e.g., multi-modal or highly anisotropic reverse trajectories). Without that comparison, it is unclear whether the control formulation’s state dependence is load-bearing or whether the method effectively reduces to learning a 1D time warp. A short ablation (feedback vs distilled) at a few NFEs would resolve this.","section":null},{"comment":"§5.2–5.3 and free parameters: The implementable scheme depends on λ (Gaussian randomization), γ (budget multiplier), ε-floor on |Q|, actor/critic architectures, and learning rates a_n. The manuscript does not report sensitivity of the learned schedule or FID to these choices, nor a protocol for selecting them across datasets. Because the headline claim is that ART-RL is a reusable, principled schedule rather than a tuned artifact, a brief sensitivity or default-protocol statement (especially for λ and the |Q| floor) is needed to support transferability claims in §6.5.","section":null}],"minor_comments":[{"comment":"Eq. (13) and the paragraph after (11): the constant bias −λ from Gaussian randomization and the compensating +λT term are clear, but a one-line display of the identity ∫(−|Q|θ²−γθ)π dθ = −|Q|μ²−γμ−λ would help readers who skip the surrounding text.","section":null},{"comment":"Figure 3 uses a log vertical axis across heterogeneous metrics (W₂, FID, LeNet-FID). Consider labeling each panel with the metric name and units, and stating in the caption that lower is better for all panels.","section":null},{"comment":"Notation: F(x,ψ) in (5) and Q(x,ψ) in (8) are dense; a short remark that Q is the second-order coefficient along the probability-flow field (not a Hessian of a loss) would reduce confusion for readers outside numerical SDEs.","section":null},{"comment":"Related work: CTRL applications to diffusion fine-tuning (Gao et al., Zhao et al.) are cited; a sentence distinguishing schedule learning (this paper) from score/policy fine-tuning would sharpen the novelty claim.","section":null},{"comment":"Table 2 NFE=35: three matched runs are mentioned in text but not tabulated with means/stds; a small parenthetical or appendix table would strengthen the claim that ART-RL still wins at the strongest EDM budget.","section":null}],"recommendation":"major_revision","confidential_remarks":"The Euler-vs-Heun/RK4 mismatch is the only load-bearing scientific gap; the theory and experimental design are otherwise careful and the transfer suite is a real strength. I would not reject on that gap alone if the authors either retrain under a higher-order residual or clearly demote the solver-agnostic language and add a mismatch ablation. Fit for a serious ML/control journal is good if the revision addresses the objective–solver alignment."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The thing worth knowing is that this is a genuine methods package, not another hand-tuned grid: they cast timestep allocation as time-reparameterization control (ART), prove that the mean of an optimal Gaussian ART-RL policy recovers the deterministic ART optimizer, and then ship actor–critic moment conditions that produce drop-in grids. Those grids beat Uniform/DPM/EDM at matched NFE while holding the score model, solver, and pipeline fixed, and the CIFAR-10 schedule transfers without retraining across budgets, AFHQv2/FFHQ/ImageNet-64, and even EDM2 latent ImageNet-512.\n\nWhat is new is the formulation plus the equivalence and implementable CTRL updates, not “RL for diffusion” in the abstract. Theorem 1 is clean under classical HJB assumptions; Theorems 2–3 give coherent policy evaluation/improvement and the martingale orthogonality that justifies the updates. Experiments are unusually careful for this claim: model and solver stay fixed, only the grid changes, and they include a 1-D analytical-score case, a small MNIST score model, Euler/Heun/RK4, and broad transfer. Distillation to a time-only schedule is pragmatic and matches the observed near-deterministic θ curves. Citations to their own CTRL line are appropriate background, not circularity.\n\nThe soft spot the stress-test flags is real and already named in §7: the objective is a leading-order Euler residual |Q|θ², yet headline image results use Heun and RK4. That is a genuine theory–practice gap; they do not derive a matching higher-order local residual. I would not over-weight it into “the gains are coincidence.” They also report Euler ablations where the same story holds, the 1-D case is pure Euler, and the transfer suite is empirical evidence that the learned allocation still helps under the deployed solvers. High-NFE margins are small (e.g. 1.82 vs 1.85 FID) but directionally consistent and repeated. Free parameters (λ, γ, ε, nets, distillation) are ordinary for actor–critic work; no code is shipped, so reproducibility is only moderate.\n\nThis is for people who care about sampling schedules, CTRL applied to generative models, or drop-in efficiency without retraining backbones. It deserves a serious referee. I would engage with it and expect revision mainly on the surrogate/solver mismatch and tighter uncertainty reporting, not a rewrite of the core argument.","headline":"Solid control-theoretic schedule learning for diffusion sampling with real transfer evidence; the Euler-surrogate vs Heun/RK4 gap is a real but already-flagged soft spot, not a collapse of the claim.","tokens_in":26442,"tokens_out":609,"would_cite":true,"duration_ms":6855,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Learning the sampling clock as continuous-time control improves diffusion sample quality by changing only the timestep grid, and the learned schedules transfer without retraining.","keywords":["generative AI","diffusion model","sampling","adaptive reparameterized time","optimal control","reinforcement learning","distillation","transfer learning"],"falsifier":"Train and distill ART-RL on CIFAR-10, plug only that grid into the official EDM Heun sampler at NFE=35 with all else fixed, and check whether FID fails to beat or match the standard EDM schedule on matched 50k-sample runs; if it does not improve, the central drop-in claim fails.","tokens_in":26148,"feed_emoji":"⏱️","tokens_out":864,"duration_ms":19568,"temperature":0.7,"pith_summary":"This paper argues that the discrete timesteps used when sampling from score-based diffusion models should be chosen by solving a continuous-time optimal control problem, not by uniform or hand-designed rules. It introduces Adaptive Reparameterized Time (ART): the sampler advances on a reparameterized clock whose speed is the control, so a uniform grid on that clock becomes an adaptive grid in physical diffusion time, with an objective based on a leading-order Euler local-error surrogate. Because the deterministic problem is high-dimensional, the authors solve an auxiliary randomized version (ART-RL) with Gaussian policies whose mean recovers the optimal ART warping rate, and they derive actor–critic updates from policy evaluation and improvement identities. Across low-dimensional analytic scores and image pipelines, swapping only the timestep list into existing samplers improves quality at matched budgets, and a schedule trained once transfers across budgets, datasets, solvers, pipelines, and representation spaces.","feed_headline":"Learned sampling clocks beat hand-crafted diffusion grids","feed_subtitle":"Change only the timestep list: quality rises at fixed budgets, and schedules transfer without retraining.","key_machinery":"Adaptive Reparameterized Time (ART): treat the speed of a reparameterized sampling clock as the control so uniform steps on that clock induce adaptive physical timesteps. ART-RL recovers the ART optimum as the mean of an optimal Gaussian policy, with implementable actor–critic updates from trajectory moment identities.","core_discovery":"Timestep allocation for diffusion sampling can be cast as a continuous-time control problem over a time-warping rate; an auxiliary Gaussian-policy RL formulation is equivalent at optimality, and the resulting learned schedules improve sample quality over Uniform, DPM, and EDM grids when plugged into unchanged samplers, with broad transfer without retraining.","pith_inferences":["The same time-warping control idea may extend to reverse-time SDE samplers, not only probability-flow ODEs, though the paper leaves that open.","If an Euler local-error proxy already helps Heun and RK4 empirically, higher-order local-error surrogates could further refine allocation for those integrators.","Collapse of learned policies to nearly time-only schedules suggests much of the gain is a one-dimensional warping rather than full state-feedback control, which would simplify deployment further."],"forward_implications":["Existing diffusion pipelines can raise sample quality at fixed evaluation budgets by replacing only the timestep schedule with an ART-RL grid.","One offline schedule-training run can be amortized across budgets, datasets, solvers, and representation spaces via distillation and transfer.","Hand-crafted schedules such as EDM and DPM can be suboptimal even on simple problems with known scores.","Schedule learning is separable from score-network training: no architectural change or solver rewrite is required."],"fun_headline_variants":["ART warps sampling clocks via continuous control and actor-critic","Learned time-warping rate beats Uniform DPM EDM grids at fixed budget","Gaussian-policy RL recovers optimal diffusion timestep schedules","Plug-in ART timesteps lift quality; schedules transfer without retraining","Control the sampling clock speed to adapt diffusion discretization"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The method chooses timesteps by minimizing a local error measure derived for simple Euler steps, even when the sampler actually uses higher-order methods and high-dimensional learned scores.","fun_headline_variants_meta":{"raw":{"variants":["ART warps sampling clocks via continuous control and actor-critic","Learned time-warping rate beats Uniform DPM EDM grids at fixed budget","Gaussian-policy RL recovers optimal diffusion timestep schedules","Plug-in ART timesteps lift quality; schedules transfer without retraining","Control the sampling clock speed to adapt diffusion discretization"]},"model":"grok-4.5","effort":"low","cost_usd":0.005656,"raw_usage":{"total_tokens":1556,"prompt_tokens":824,"num_sources_used":0,"completion_tokens":73,"cost_in_usd_ticks":56560000,"prompt_tokens_details":{"text_tokens":824,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":659,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":824,"tokens_out":73,"duration_ms":7168,"temperature":1.0,"reasoning_tokens":659,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T08:24:29.769439+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train and distill ART-RL on CIFAR-10, plug only that grid into the official EDM Heun sampler at NFE=35 with all else fixed, and check whether FID fails to beat or match the standard EDM schedule on matched 50k-sample runs; if it does not improve, the central drop-in claim fails.","supporting_citations":[],"review_version":2}