{"id":"a7641208-f64d-43be-99c0-611c7a9e0dd4","arxiv_id":"2502.00688","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"HOMO extends shortcut diffusion with acceleration and jerk supervision, but the proof of superior approximation is not supported and experiments lack error bars.","lead":"This paper adds acceleration and higher-order derivative terms to one-step shortcut diffusion models, claiming smoother and more accurate generation. The theoretical support is not established: the key error bound reduces to the first-order result plus an uncontrolled extra term.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5.1's bound has an uncontrolled E[||v_t - a_t||^2] term, making it strictly looser than the first-order bound, so it does not support the claim that HOMO outperforms first-order methods.","rationale":"The reader's weakest assumption identifies the uncontrolled E[||v_t - a_t||^2] term, and I agree this is the most load-bearing concern. The theorem's bound is not just unhelpful for proving superiority; it is strictly worse than the first-order bound from Lemma C.18, so the theoretical argument is internally inconsistent with the claimed conclusion. The schedule mismatch with Assumption C.15 and the absence of the self-consistency loss from the theory further weaken the connection between theory and experiments. The paper does present a plausible empirical idea and some qualitative results, but the central theoretical claim fails, and the empirical evidence lacks error bars, seeds, and code. Therefore the REJECT verdict remains appropriate.","tokens_in":40893,"tokens_out":4472,"duration_ms":40375,"concrete_test":"For the VP schedule α_t = exp(-a(1-t)^2/4 - b(1-t)/2) with a=19.9, b=0.1, compute E[||\\dot{x}_t - \\ddot{x}_t||^2] analytically at t = 0, 0.25, 0.5, 0.75, 1 using x0~N(0,I) and x1~data, and compare it with (\\dotα_t^2 log N + \\dotβ_t^2) N^{-2s/d} for a representative N (e.g., 10^4) and dimension d=2. If the uncontrolled term is not negligible relative to the first-order term, then Theorem 5.1's bound is vacuous and cannot establish that HOMO outperforms first-order methods.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that high-order supervision ensures superior approximation accuracy is not established by Theorems 5.1 and 5.2. Both theorems bound the error by the first-order approximation error (\\dotα_t^2 log N + \\dotβ_t^2) N^{-2s/d} plus an additive term E_{x~P_t}[||\\dot{x}_t^{true} - \\ddot{x}_t^{true}||^2], which is never controlled. Since Lemma C.18 already provides the first-order bound, the theorem's RHS is strictly larger than the first-order bound by this positive term, so no superiority follows. The proof (D.1) simply bounds the acceleration error by the velocity approximation error plus the distance between velocity and acceleration, not by any second-order structure. For the experimental VP schedule α_t = exp(-a(1-t)^2/4 - b(1-t)/2) with a=19.9, b=0.1, at t=0 the uncontrolled term is approximately 0.28·d (since \\dotα(0)≈0.066 and \\ddotα(0)≈0.594), which does not vanish as N grows, while the first-order term can be made arbitrarily small. Moreover, Assumption C.15 requires α_t ~ b0 t^κ near t=0, but the experimental schedule has α_0 = exp(-a/4 - b/2) ≈ 0.0066 > 0, so the theoretical regime does not contain the experiments. The self-consistency loss in Definition 4.3, central to the algorithm, is also absent from the theoretical analysis. Thus the paper's stated theoretical guarantee does not support its headline claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HOMO, an extension of one-step Shortcut diffusion models that adds higher-order trajectory supervision (acceleration, jerk, and beyond) to the training objective and uses a second-order Taylor-style update at inference. The main theoretical claims are Theorems 5.1 and 5.2 (and their formal versions in Appendix D), which bound the combined velocity and acceleration approximation error, and the paper argues that this establishes superiority over first-order methods. Experiments on synthetic Gaussian-mixture and spiral datasets compare ablations of the first-order, second-order, and self-consistency losses, and a third-order variant is also tested.","tokens_in":41207,"tokens_out":5812,"duration_ms":55980,"significance":"If the claimed superiority were actually established, high-order supervision for shortcut diffusion models could be a meaningful contribution to one-step generative modeling. The paper does provide a clear algorithmic framework and a fairly extensive set of synthetic ablations. However, the central theoretical result does not deliver the claimed high-order advantage: it reduces to the first-order bound from prior work plus an uncontrolled additive term. The experimental schedule used in Section 6.1 also falls outside the assumptions required by the theory. The self-consistency loss, which is central to the algorithm, is absent from the theoretical analysis. The empirical evaluation is limited to low-dimensional synthetic distributions, so the broader claims about vision generation are not supported by the presented experiments. Overall, the paper's main claim is not established by its own theorems.","major_comments":[{"comment":"The proof of Theorem 5.1 bounds the acceleration approximation error by applying the triangle inequality to the first-order approximation error plus the term E_{x~P_t}[||\\dot{x}_t^{true} - \\ddot{x}_t^{true}||^2]. Since this additive term is nonnegative and is never controlled, the theorem's right-hand side is not smaller than the first-order bound of Lemma C.18; it is the first-order bound plus an uncontrolled bias. Consequently, the abstract's claim that \"HOMO's high-order supervision ensures superior approximation accuracy, outperforming first-order methods\" does not follow from the theorem. The same issue propagates through Theorems D.2 and D.4 by induction, so the higher-order guarantees inherit the same limitation.","section":"Theorem 5.1 / Appendix D.1, Eq. (3)"},{"comment":"Assumption C.15 requires alpha_t = b0 t^kappa for sufficiently small t, which implies alpha_t tends to 0 as t tends to 0. The experiments in Section 6.1 use the VP schedule alpha_t = exp(-a(1-t)^2/4 - b(1-t)/2) with a=19.9 and b=0.1, for which alpha_0 = exp(-5.025) approximately 0.0066, which is strictly positive. The theoretical regime therefore does not contain the experimental setting, and Theorems 5.1 and 5.2 do not apply to the reported experiments. Moreover, for this schedule the uncontrolled term E[||\\dot{x}_t^{true} - \\ddot{x}_t^{true}||^2] does not vanish as N grows: at t=0, |\\dot{alpha}(0) - \\ddot{alpha}(0)| is approximately 0.525, so this term is on the order of the dimension d, while the first-order term is O(N^{-2s/d}).","section":"Assumption C.15 vs Section 6.1"},{"comment":"The self-consistency target in Definition 4.2 is defined as the average of the network's own outputs u_{1,theta1}(x_t, t, d) and u_{1,theta1}(x_{t+d}, t, d). As written, the loss in Definition 4.3 uses this target without any stop-gradient, so the objective is a moving target that depends on the same parameters being trained; such a loss can be trivially minimized by a constant prediction. Algorithm 1 applies stopgrad, but the formal definitions do not state this, and the theoretical analysis in Section 5 and Appendix D does not analyze the self-consistency loss at all. The method whose properties are proved is therefore not exactly the method whose empirical performance is reported.","section":"Definitions 4.2 and 4.3 vs Algorithm 1"},{"comment":"Theorems 5.1 and 5.2 are presented as \"informal versions\" with \"some mild assumptions,\" yet the abstract and conclusion treat them as establishing a definitive superiority guarantee. The formal statements in Appendix D require a specific set of assumptions (Besov smoothness of p0, support conditions, the alpha_t form of Assumption C.15, and Lipschitz conditions) that are not verified for the experimental Gaussian-mixture or spiral datasets. Given that the informal theorem statements omit these conditions, the current presentation overstates the strength and scope of the theoretical results.","section":"Section 5 and abstract"}],"minor_comments":[{"comment":"The abstract and introduction use promotional language such as \"game-changing,\" \"revolutionary leap,\" \"unprecedented,\" and \"shatters these barriers,\" which is inappropriate for a scientific paper and should be replaced with neutral, precise claims.","section":"Abstract and introduction"},{"comment":"Section 6.1 states that M1+SC is equal to the original One-step Shortcut model, but Algorithm 3 in Appendix A defines the Shortcut model with the linear interpolation x_t = (1-t)x_0 + t x_1, whereas the HOMO experiments use the VP ODE schedule with alpha_t and beta_t. The equivalence is therefore not immediate and should be clarified.","section":"Section 6.1 and Appendix A"},{"comment":"Theorem 5.1 says \"Let N be a value associated with sample size n\" without defining the relationship; Remark C.10 later sets N = n^{d/(2s+d)}, but this should be stated in the main text where the theorem is used.","section":"Theorem 5.1 and Remark C.10"},{"comment":"There are typographical errors in the proofs: in D.1 the expectation E_{x~P_t}[||\\dot{x}_t^{true} - \\ddot{x}_t^{true}||^2] is missing a closing bracket, and in D.3 the term \"||2 \\dot{x}_t^{true} - \\ddot{x}_t^{true}||\" appears where the intended expression is \"||\\dot{x}_t^{true} - \\ddot{x}_t^{true}||.\" These should be corrected.","section":"Appendix D.1 and D.3"},{"comment":"The computational cost table lists M1+M2 as 16.960 MFLOPS but M1+M2+SC as 68.480 MFLOPS, the same value as M2+SC. It is unclear why adding the M1 term to M2+SC does not increase the FLOP count; this inconsistency should be explained.","section":"Table 5"},{"comment":"The related-work sections contain many citations to large-language-model and foundation-model papers that are not connected to the present topic of shortcut diffusion or higher-order flow matching; this material should be removed or substantially trimmed.","section":"Section 2 and Appendix B"}],"recommendation":"reject","confidential_remarks":"The central theoretical claim is not supported by the theorems as written: the bounds in Theorems 5.1 and 5.2 reduce to the first-order result of Lemma C.18 plus an uncontrolled additive term, so no high-order superiority follows. The mismatch between Assumption C.15 and the experimental VP schedule is severe, and the self-consistency loss is not part of the theoretical analysis. These are load-bearing issues rather than presentational ones. The paper also contains a large number of citations that appear irrelevant to the topic, and the empirical evaluation is restricted to small synthetic datasets despite the abstract's claims about vision generation. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague, the key thing you should know: HOMO's headline claim, that high-order supervision provably gives superior approximation accuracy, is not supported by the paper's own theorems. The bounds in Theorem 5.1 and 5.2 are the old first-order bounds plus an extra E[||v_t - a_t||^2] term that is never controlled. Since that term is positive, the bound is strictly looser than the first-order bound. On top of that, the experimental VP schedule violates the assumption α_t ~ t^κ near t=0, so the theory does not even cover the experiments.\n\nWhat is genuinely good: the idea of adding acceleration and jerk supervision to shortcut models is natural and reasonably motivated. The training and sampling algorithms are clearly described. The ablation study over M1, M2, SC and their combinations is thorough, and on the synthetic 2D datasets the full model M1+M2+SC does consistently beat the first-order+SC baseline. That is a modest but real empirical result.\n\nThe soft spots are significant. The self-consistency loss, a bootstrapped target from the network's own predictions, is central to the algorithm but appears in none of the theorems. The experiments report single numbers per configuration, with no error bars, no seeds, and no code. Appendix H contains an obvious contradiction: it says the Shortcut model (M1+SC) is 33.6% faster than M1, but the table shows 283 vs 477 it/s, which is slower. That kind of internal inconsistency makes the computational analysis unreliable.\n\nSo who should read it? If you work on shortcut models or high-order flow matching, the empirical ablations are worth a skim, and the idea could spark something. I would not cite it as evidence that high-order supervision is provably better, because that proof is not there.\n\nFor peer review: I would send it to a serious referee rather than desk-rejecting. The idea is relevant, the experiments, though flawed, are reproducible in principle, and the theoretical gap is exactly what a good referee could help the authors fix. I would expect major revision, and if the uncontrolled term and the schedule mismatch are not addressed, the central claim should be withdrawn.\n\nFinal recommendation: treat it as a promising but not-yet-supported contribution. If you are an editor, give it a chance with clear instructions to the referee about the E[||v_t - a_t||^2] term.","headline":"The theoretical claim of superior approximation is not proven because the bounds are strictly looser than first-order, but the empirical idea has merit and the paper deserves a serious referee.","tokens_in":41803,"tokens_out":2961,"would_cite":false,"duration_ms":29726,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","41A25","65D07"],"pacs":[],"model":"deepseek-v4-flash","headline":"HOMO claims that adding acceleration and jerk supervision to one-step shortcut diffusion yields provably better trajectory approximation and empirically sharper matches of high-curvature distributions than the velocity-only Shortcut model.","keywords":["high-order flow matching","shortcut diffusion","one-step generative models","acceleration supervision","self-consistency target","trajectory approximation","Besov space","generative transport"],"falsifier":"Measure $\\mathbb{E}_{x\\sim P_t}[\\|\\dot x_t^{\\mathrm{true}}-\\ddot x_t^{\\mathrm{true}}\\|^2]$ under the paper's VP schedule $\\alpha_t=\\exp(-19.9(1-t)^2/4-0.1(1-t)/2)$; if this residual is comparable to or larger than the claimed $N^{-2s/d}$ rate, the theorems' additive term dominates and the approximation guarantee is vacuous. A second decisive check is a seeded re-run of the eight-mode Gaussian experiment comparing M1+SC with M1+M2+SC: if the Euclidean-distance losses are statistically tied, the central empirical claim fails.","tokens_in":40600,"feed_emoji":"⚡","tokens_out":7810,"duration_ms":76737,"temperature":0.7,"pith_summary":"One-step shortcut diffusion models learn a velocity field that transports noise to data in very few steps, but the velocity-only objective is poorly matched to distributions with sharp bends. This paper proposes HOMO, which adds acceleration (and, in an extension, jerk) to both the training loss and the sampling update, turning each step into a second-order Taylor step. The claim is that high-order supervision yields provably smaller trajectory approximation error and, in practice, noticeably better matches of multimodal and high-curvature synthetic targets. The authors prove approximation bounds using Besov-space B-spline network approximation and verify the improvement on Gaussian mixtures and spiral/circle datasets. If correct, the work offers a simple recipe—supervise time-derivatives beyond velocity—to improve few-step generative transport.","feed_headline":"High-order supervision sharpens one-step shortcut diffusion","feed_subtitle":"Matching acceleration in addition to velocity yields smoother, more accurate trajectories in one-shot generative transport.","key_machinery":"The load-bearing mechanism is high-order Taylor supervision of the transport trajectory. Sampling advances by $x_{t+d}=x_t+d\\,u_1(x_t,t,d)+\\frac{d^2}{2}u_2(u_1(x_t,t,d),x_t,t,d)$, so the second network $u_2$ corrects curvature rather than only slope. Training minimizes three terms: velocity matching $\\|u_1-\\dot x_t^{\\mathrm{true}}\\|^2$, acceleration matching $\\|u_2-\\ddot x_t^{\\mathrm{true}}\\|^2$, and a self-consistency loss that averages $u_1$ at $x_t$ and $x_{t+d}$ to stabilize long-horizon prediction. The proofs approximate the true velocity and acceleration fields by B-spline-based ReLU networks in Besov spaces, transferring known approximation rates; an induction step extends the argument from second order to arbitrary fixed order $K$, at the cost of an uncontrolled residual term measuring the gap between consecutive time-derivatives.","core_discovery":"The central claim is that for the interpolation $x_t = \\alpha_t x_0 + \\beta_t x_1$ used by Shortcut models, supervising a network pair $(\\phi_1,\\phi_2)$ against both the true velocity $\\dot x_t$ and true acceleration $\\ddot x_t$ gives an approximation bound of the form $\\int (\\|\\phi_1-\\dot x_t\\|^2+\\|\\phi_2-\\ddot x_t\\|^2)p_t(x)\\,dx \\lesssim (\\dot\\alpha_t^2\\log N + \\dot\\beta_t^2)N^{-2s/d} + \\mathbb{E}_{x\\sim P_t}[\\|\\dot x_t-\\ddot x_t\\|^2]$, with an analogous large-$t$ bound replacing $N^{-2s/d}$ by $N^{-\\eta}$. The training objective combines velocity matching, acceleration matching, and a self-consistency term that averages the model's velocity at two nearby states; sampling iterates $x_{t+d}=x_t+d\\,u_1+\\frac{d^2}{2}u_2$. Empirically, the combination of first-order, second-order, and self-consistency losses beats every ablation and the Shortcut baseline on the reported Euclidean-distance losses, and third-order HOMO continues that trend on the tested synthetic distributions.","pith_inferences":["Editorial inference: the same second-order Taylor loss could be grafted onto other flow-matching or rectified-flow objectives, not just Shortcut models, since the theoretical framework is written for the general field $x_t=\\alpha_t x_0+\\beta_t x_1$.","Editorial inference: because the residual term in the theorems is uncontrolled, the clearest way to test the proof's regime is to measure $\\mathbb{E}_{x\\sim P_t}[\\|\\dot x_t-\\ddot x_t\\|^2]$ under the paper's actual VP schedule; the paper does not report this quantity.","Editorial inference: the empirical evidence is limited to low-dimensional synthetic distributions, so whether the benefit transfers to high-dimensional image or video generation remains an open extension.","Editorial inference: the claim that high-order supervision outperforms first-order supervision is about matching a chosen analytic interpolation; it does not by itself establish that the learned transport map matches the true data-generating distribution, since the acceleration target inherits the coupling $(x_0,x_1)$ used to define the trajectory."],"forward_implications":["For the same number of sampling steps, HOMO's update includes an explicit curvature correction, so trajectories should follow the underlying transport ODE more closely than the velocity-only Shortcut update.","If the ablations generalize, the second-order loss and the self-consistency loss carry different information: self-consistency sharpens the learned distribution, while acceleration matching supplies the geometric detail needed for outer boundaries and spiral arms.","The theoretical bounds imply that the advantage of high-order supervision is strongest when the velocity field changes rapidly but the gap between velocity and acceleration stays small relative to the $N^{-2s/d}$ rate.","Third-order HOMO extends the same Taylor correction to jerk and shows continued quantitative improvement on the reported spin and dot-circle benchmarks.","The computational-cost section indicates that the added accuracy comes with a measurable but not prohibitive increase in FLOPs and parameters relative to the base Shortcut configuration."],"supporting_citations":[{"why":"Defines the Shortcut model and its first-order self-consistency training; HOMO's M1+SC configuration is exactly that baseline, and the paper compares against it throughout.","marker":"[FHLA25]"},{"why":"Supplies the Besov-space and B-spline approximation lemmas (Lemma C.18 and Lemma C.19) that Theorems 5.1 and 5.2 directly reuse to bound the network approximation error.","marker":"[FSI+24]"},{"why":"Provides the VP ODE interpolation $x_t=\\alpha_t x_0+\\beta_t x_1$ and the noise schedule used in all of the paper's experiments.","marker":"[LGL22b]"},{"why":"Establishes near-optimal deep ReLU approximation rates in Besov spaces, which underlies the B-spline network approximation used in the proofs.","marker":"[Suz19]"},{"why":"Contributes the diffusion-model minimax estimation perspective and Besov-space approximation tools that the proof framework depends on.","marker":"[OAS23]"},{"why":"Provides the definition and standard properties of Besov spaces on which Assumption C.13 and the smoothness framework rest.","marker":"[Tri92]"}],"fun_headline_variants":["High-order matching sharpens one-step diffusion","Acceleration supervision yields smoother diffusion trajectories","Beyond velocity: high-order supervision for diffusion","HOMO: high-order matching boosts one-step generative models","High-order supervision stabilizes one-step shortcut diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof never bounds the additive term $\\mathbb{E}_{x\\sim P_t}[\\|\\dot x_t-\\ddot x_t\\|^2]$ left in Theorems 5.1 and 5.2, and the paper's experimental schedule $\\alpha_t=\\exp(-19.9(1-t)^2/4-0.1(1-t)/2)$ lies outside Assumption C.15's $\\alpha_t\\sim b_0 t^\\kappa$ regime.","fun_headline_variants_meta":{"raw":{"variants":["High-order matching sharpens one-step diffusion","Acceleration supervision yields smoother diffusion trajectories","Beyond velocity: high-order supervision for diffusion","HOMO: high-order matching boosts one-step generative models","High-order supervision stabilizes one-step shortcut diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1477,"prompt_tokens":1028,"completion_tokens":449,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":380}},"tokens_in":644,"tokens_out":449,"duration_ms":4455,"temperature":1.0,"reasoning_tokens":380,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T18:05:26.482111+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure $\\mathbb{E}_{x\\sim P_t}[\\|\\dot x_t^{\\mathrm{true}}-\\ddot x_t^{\\mathrm{true}}\\|^2]$ under the paper's VP schedule $\\alpha_t=\\exp(-19.9(1-t)^2/4-0.1(1-t)/2)$; if this residual is comparable to or larger than the claimed $N^{-2s/d}$ rate, the theorems' additive term dominates and the approximation guarantee is vacuous. A second decisive check is a seeded re-run of the eight-mode Gaussian experiment comparing M1+SC with M1+M2+SC: if the Euclidean-distance losses are statistically tied, the central empirical claim fails.","supporting_citations":[],"review_version":1}