{"id":"0681e816-1fa5-43ca-8466-5a56b3b43fcd","arxiv_id":"2508.17426","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"Modular MeanFlow interpolates between stop-gradient and full-gradient training for one-step generative models via a warmup schedule, reporting a CIFAR-10 FID of 3.41.","lead":"This paper introduces Modular MeanFlow, a training objective for one-step generative models that blends a MeanFlow-style velocity loss with a schedule that gradually increases gradient information. The authors report improved sample quality and robustness on image and trajectory tasks, but the theoretical derivation contains a scaling error and no code or error bars are provided.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Loss target (x1−x0)/(t−r) contradicts the §5.1 interpolation, whose average velocity is (x1−x0)/(1−r); every sampled t<1 is trained against a mis-scaled vector.","rationale":"The reader's weakest_assumption is precisely the load-bearing flaw. I re-derived Eqs. (8)–(11) from the interpolation given in §5.1. The paper's target is off by a factor (1−r)/(t−r) relative to the exact average velocity. This matters for every training sample with t<1 (i.e., almost all samples), so the central loss does not implement the MeanFlow identity it claims. The curriculum and gradient-modulation comparisons in §5.2–5.4 therefore compare variants of a mis-specified objective; the reported FID 3.41 etc. cannot validate the proposed MMF formulation. I do not see an alternative reading of the setup that rescues the target: either the interpolation definition must change or the target must be (x1−x0)/(1−r). The paper has no code, no error bars, and the CM baseline is missing, but the derivation gap is sufficient for rejection. I concur with the reader's REJECT verdict; no adjustment is needed.","tokens_in":10396,"tokens_out":7350,"duration_ms":69408,"concrete_test":"Let x0=(1,0), x1=(0,1), r=0.25, t=0.75. §5.1 gives xt=(0.5,0.5); exact average velocity (xt−x0)/(t−r)=(-1,1), corrected target (x1−x0)/(1−r)=(-1.333,1.333), paper target (x1−x0)/(t−r)=(-2,2). A direct symbolic check: substitute the §5.1 interpolation into Eq. (11); since u=(x1−x0)/(1−r) and d/dt u=0, the left side equals (x1−x0)/(1−r), not the paper's right side. As a computational confirmation, retrain the smallest CIFAR-10 variant with the corrected target (x1−x0)/(1−r) for 100k steps and compare loss and FID against the paper's target; material differences confirm the mis-scaling.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 replaces x1−x0 by (t−r)u and derives the regression target (x1−x0)/(t−r) in Eqs. (8)–(11). Section 5.1 defines the training path as xt = (1−α)x0 + αx1 with α = (t−r)/(1−r). For this path, x_t − x_r = ((t−r)/(1−r))(x1−x0), so the exact average velocity is u = (x_t − x_r)/(t−r) = (x1−x0)/(1−r). The paper's target differs by the factor (1−r)/(t−r), matching only at t = 1. Because (r,t) are sampled with r < t, essentially every training example is supervised against a mis-scaled vector field. The total derivative term in Eq. (11) does not repair this: along the linear path u is constant in t, so the left side evaluates to u, not (x1−x0)/(t−r). Consequently the loss Lλ is not the MeanFlow-inspired objective claimed, and the reported FID, one-step error, and generalization results cannot be attributed to the stated method. The paper also omits code, error bars, and the Consistency Model baseline, but the scaling inconsistency is sufficient to invalidate the central derivation as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces Modular MeanFlow (MMF), a training objective for one-step generative modeling that combines a MeanFlow-inspired regression target with a stop-gradient interpolation parameter λ and a curriculum schedule over λ. The authors claim that this loss family unifies consistency and flow-matching objectives, avoids higher-order derivatives, and achieves FID 3.41 on CIFAR-10 with one-step sampling, with additional results on few-shot and out-of-distribution settings, ODE fitting, and control. The theoretical derivation starts from an identity linking instantaneous and average velocity and proposes the loss Lλ in Eq. (11); the empirical section compares λ=0, λ=0.5, λ=1, and curriculum variants against MeanFlow baselines.","tokens_in":10679,"tokens_out":5551,"duration_ms":50682,"significance":"The central idea of tuning gradient flow through a stop-gradient interpolation is practically relevant, and a well-supported method of this kind could be useful for one-step generation. However, as written the paper does not deliver a machine-checked proof, reproducible code, or a parameter-free derivation, and the main loss is derived from an inconsistent velocity surrogate. The reported FID and error numbers therefore cannot be attributed to the stated method. The empirical study also lacks error bars, a Consistency Model baseline, and any ImageNet-64 result promised in Section 5.1. If the derivation were corrected, the modular loss idea would still need a substantial experimental re-run; in its current form the manuscript does not establish its central claims.","major_comments":[{"comment":"The regression target is (x1-x0)/(t-r), but the average velocity of the training path defined in Section 5.1 is (x1-x0)/(1-r). Specifically, xt=(1-α)x0+αx1 with α=(t-r)/(1-r) gives x_t-x_r=((t-r)/(1-r))(x1-x0), hence u=(x_t-x_r)/(t-r)=(x1-x0)/(1-r). The target differs by the factor (1-r)/(t-r) and matches only at t=1. Since r<t is sampled with probability one, essentially all training examples are supervised against a mis-scaled vector. The total-derivative term in Eq. (11) does not repair this because u is constant along the linear path and its total derivative is zero. This invalidates the claimed MeanFlow objective and the attribution of the reported FID, one-step error, and generalization results.","section":"§4.2, Eq. (8)-(11); §5.1"},{"comment":"The assertion that Lλ unifies and generalizes full MeanFlow, StopGrad MeanFlow, and Consistency Models is stated without derivation. Table 1 gives a qualitative comparison, but no formal statement or proof shows how specific choices of λ, r, and t recover those objectives. This matters because the claimed theoretical grounding rests on that unification.","section":"§4.5, Table 1"},{"comment":"The experimental section promises ImageNet-64 and the Consistency Model baseline in Section 5.1 but never reports results for either; no standard errors or confidence intervals are given for FID or one-step error; and no code or checkpoints are provided. Given that the main contribution is a training objective, the absence of these details makes the empirical superiority claims unverifiable.","section":"§5.1, §5.2"}],"minor_comments":[{"comment":"The schedule λ(ttrain) uses ttrain, but the time variable t is already used for the path in Eq. (2); please clarify the notation to avoid confusion.","section":"§4.4, Eq. (12)"},{"comment":"The interpolation parametrization with α=(t-r)/(1-r) makes x_r=x0 and x_t=x1, while the ODE in Eq. (1) is written with x1 as the prior and x0 as data; the direction of the path should be stated explicitly.","section":"§5.1"},{"comment":"Several references appear unrelated to the statements they are attached to (e.g., [11] for the harmonic potential in Section 5.6, and [22]-[23] for consistency formulations in Section 1); please recheck.","section":"References"},{"comment":"The claim that the JVP overhead is 'typically <20% extra time' is not accompanied by any measurement or protocol.","section":"§4.5"},{"comment":"The bar charts and box plots give no error bars or sample sizes, and the text describing 'variance' in Figure 4b does not define what variance is plotted.","section":"Figures 3 and 4"}],"recommendation":"reject","confidential_remarks":"Most of the load-bearing issues are in the manuscript itself, so the recommendation is based on the derivation inconsistency. Separately, the citation list contains several entries that do not appear to support the statements they are attached to; this is worth editorial scrutiny but is not part of the technical verdict. There is no code release or reproducibility artifact, which is a further concern for an empirical paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core of the paper is a nice idea: interpolate between stop-gradient and full backprop through the average-velocity objective with a warmup schedule for λ. That is clean and worth exploring. The low-data and distribution-shift experiments are also a thoughtful addition, and Table 1 is a genuinely useful comparison of training regimes.\n\nThe problem is the derivation. In Section 4.2, the paper substitutes x1−x0≈(t−r)u into the MeanFlow identity and ends up with a regression target (x1−x0)/(t−r). But Section 5.1 defines the interpolation as x_t=(1−α)x0+αx1 with α=(t−r)/(1−r). For that path, x_r=x0 and x_t−x_r=((t−r)/(1−r))(x1−x0), so the exact average velocity is (x1−x0)/(1−r), not (x1−x0)/(t−r). The two match only when r=0. Since r<t≤1 in training, essentially every sampled target is mis-scaled. This is not a cosmetic typo; it changes the loss landscape pointwise and means the reported FID, one-step error, and generalization numbers do not test the stated objective.\n\nThere are other soft spots, but they are minor by comparison: no code, no error bars, and the Consistency Model baseline is listed but never reported in Table 2. The reference list also looks padded—self-citations appear in contexts that don't support their claims, and at least one (the hydrochory paper) is clearly irrelevant to the point being made.\n\nThe paper is not a waste of time. The λ-curriculum is a reasonable mechanism, and the paper is honest enough about the stability motivation. But the load-bearing math is wrong as written. If the target scaling is corrected and the experiments rerun, the method might still work or even improve; we can't know from this version. The authors should also add error bars and actually compare against the consistency baseline.\n\nWho is this for? Researchers working on one-step generative training objectives, especially those trying to reconcile stability and expressiveness. They would get a useful idea and a cautionary example.\n\nRecommendation: send it to peer review, but the reviewers should be told to check the target scaling carefully. As written, the central claim fails; with a corrected derivation and new experiments, it could become a solid incremental contribution.","headline":"The central loss derivation is mis-scaled, and the experiments as written don't test the claimed method; the underlying scheduling idea is plausible but needs a corrected derivation and rerun.","tokens_in":11202,"tokens_out":2509,"would_cite":false,"duration_ms":27078,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a modular loss with annealed gradient flow enables stable, one-step generative training that reaches FID 3.41 on CIFAR-10.","keywords":["one-step generation","mean flow","average velocity","gradient modulation","curriculum learning","consistency models","flow matching","generative modeling"],"falsifier":"Train the same architecture on CIFAR-10 for the same 800,000 steps with the target $(x_1-x_0)/(1-r)$ in place of $(x_1-x_0)/(t-r)$ while keeping the curriculum. If the average-velocity interpretation is correct, the corrected scaling should not hurt FID and should reduce one-step error; if the current scaling is the right one, the corrected loss should degrade.","tokens_in":10125,"feed_emoji":"⚡","tokens_out":5759,"duration_ms":53371,"temperature":0.7,"pith_summary":"This paper argues that one-step generative modeling can be made both stable and expressive by learning a time-averaged velocity field with a tunable loss. Instead of committing to either full backpropagation or stop-gradient training, Modular MeanFlow anneals the gradient flow from coarse supervision to fully differentiable training. The claimed payoff is competitive sample quality, FID 3.41 on CIFAR-10 with a one-step error of 0.076, together with stable convergence and better generalization in low-data and out-of-distribution settings. If right, this gives a practical recipe for fast generation that avoids expensive higher-order derivatives.","feed_headline":"Annealed gradients give one-step models stable training and FID 3.41","feed_subtitle":"Modular MeanFlow learns a time-averaged velocity field with a curriculum that balances stability and expressiveness.","key_machinery":"The carrying object is the modular loss $L_\\lambda$ built on the stop-gradient modulation $SG_\\lambda[z]=\\lambda z+(1-\\lambda)\\text{stopgrad}(z)$, together with the warm-up schedule $\\lambda(t_{\\text{train}})=\\min(1,t_{\\text{train}}/T_{\\text{warmup}})$. This interpolation controls how much of the Jacobian-vector product $\\nabla_x u_\\theta \\cdot v$ is allowed to backpropagate: full gradient flow at $\\lambda=1$, no gradient flow at $\\lambda=0$, and a smooth schedule in between. The loss is derived from the MeanFlow identity so that it supplies a time-averaged velocity target, and the modulation is what makes training stable without sacrificing expressiveness.","core_discovery":"The central claim is that the MeanFlow differential identity, $v(x_t,t)=u(x_t,r,t)+(t-r)\\frac{d}{dt}u(x_t,r,t)$, can be turned into a family of regression losses in which the network's average velocity and its local dynamics are compared with the displacement $(x_1-x_0)/(t-r)$, while gradient flow through the second-order term is controlled by a parameter $\\lambda$. With $\\lambda=0$ the loss is stop-gradient and stable; with $\\lambda=1$ it is fully backpropagated and expressive; the curriculum $\\lambda(t)=\\min(1,t/T_{\\text{warmup}})$ moves smoothly from one to the other. The paper claims this curriculum regime achieves the best balance, producing FID 3.41 and one-step error 0.076 on CIFAR-10, and that the framework subsumes consistency-model and flow-matching objectives as special cases.","pith_inferences":["Beyond what the paper tests, the residual one-step error could be probed by replacing the target $(x_1-x_0)/(t-r)$ with the exact average velocity $(x_1-x_0)/(1-r)$ for the linear paths actually used in training.","The paper anneals a single scalar $\\lambda$; an extension would be to anneal the velocity-matching term and the Jacobian term at different rates, which might further stabilize the later training phase.","If the loss family truly generalizes consistency models, then consistency-training techniques such as boundary conditioning or exponential moving averages could be imported into Modular MeanFlow to reduce one-step error further."],"forward_implications":["One-step sampling reduces to $x_0=x_1-u_\\theta(x_1,0,1)$, so generation needs a single function evaluation.","The same training objective transfers beyond images to tasks such as ODE trajectory fitting and control-path synthesis.","The curriculum schedule gives a principled way to choose between stability and expressiveness, and it can be tuned per domain.","Because the framework unifies consistency-based and flow-matching losses, improvements to the curriculum or to the modulation should transfer to those methods.","The low-data and distribution-shift results suggest that one-step velocity learning can be applied where full diffusion training would overfit."],"supporting_citations":[{"why":"Supplies the MeanFlow average-velocity formulation and the differential identity on which MMF's loss is built, plus the full and stop-grad baselines.","marker":"[9]"},{"why":"Defines consistency models, the concurrent one-step method that MMF's modular objective generalizes and compares against.","marker":"[35]"},{"why":"Gives flow matching, whose velocity regression target MMF reinterprets through time-averaged velocities.","marker":"[27]"},{"why":"Provides the FID metric used to measure the paper's central sample-quality claims.","marker":"[13]"},{"why":"Provides the CIFAR-10 benchmark on which the main comparison and low-data experiments are run.","marker":"[20]"}],"fun_headline_variants":["Curriculum gradient control stabilizes one-step generative models","One-step flow hits FID 3.41 via annealed gradient schedule","Modular MeanFlow: stable one-step training without higher-order terms","Time-averaged velocity fields unify consistency and flow matching","Annealed gradients enable robust one-step generation at low data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The derivation assumes that the finite difference $(x_1-x_0)/(t-r)$ equals the true average velocity over the interval, while for the linear interpolation used in training the true average is $(x_1-x_0)/(1-r)$; these coincide only at $r=0$.","fun_headline_variants_meta":{"raw":{"variants":["Curriculum gradient control stabilizes one-step generative models","One-step flow hits FID 3.41 via annealed gradient schedule","Modular MeanFlow: stable one-step training without higher-order terms","Time-averaged velocity fields unify consistency and flow matching","Annealed gradients enable robust one-step generation at low data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0003,"raw_usage":{"total_tokens":1701,"prompt_tokens":882,"completion_tokens":819,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":733}},"tokens_in":498,"tokens_out":819,"duration_ms":8147,"temperature":1.0,"reasoning_tokens":733,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:04:56.208391+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same architecture on CIFAR-10 for the same 800,000 steps with the target $(x_1-x_0)/(1-r)$ in place of $(x_1-x_0)/(t-r)$ while keeping the curriculum. If the average-velocity interpretation is correct, the corrected scaling should not hurt FID and should reduce one-step error; if the current scaling is the right one, the corrected loss should degrade.","supporting_citations":[],"review_version":2}