{"id":"e12fde0a-6b6d-4782-a073-696c89c2b513","arxiv_id":"2412.03487","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Discrete flow matching with arbitrary probability paths is made tractable via kinetic-optimal velocities, mixture schedulers, and a new ELBO, improving text, crystal, and image generation.","lead":"This paper gives discrete generative models the freedom to use any user-defined corruption path, not just masking, by providing closed-form velocity formulas and an optimality argument that selects mixture paths. The authors demonstrate the expanded design space on text, crystal, and image generation, where their kinetic-optimal and metric-based paths frequently beat masked and autoregressive baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The zero-probability extension of the kinetic-optimal flux is not rigorously justified: with w_t=1/p_t, any flux into a zero-probability state has infinite energy, so Eq. 26 is proven optimal only for p_t>0, not for arbitrary support-changing paths.","rationale":"The reader's weakest assumption correctly identifies two related weaknesses: the kinetic energy objective is chosen for closed-form solvability, and the zero-probability extension rests on a limiting argument. My stress test sharpens the second point: the proof of Proposition B.1 requires p_t>0, and with w_t=1/p_t the energy is not even finite for fluxes into zero-probability states that are gaining mass. This is a real technical gap in the over-broad statement that Eq. 26 is the kinetic-optimal flux for any discrete path. However, it does not invalidate the main practical contribution: for the mixture paths and metric paths actually used, p_t is positive on (0,1), so Eq. 26 is a valid generating velocity, and the closed-form scheduler and ELBO constructions remain coherent. The empirical claims are plausible but still need released code and repeated runs; the reader's CONDITIONAL verdict is appropriate. My concrete test would settle whether the optimality claim needs qualification, and my verdict remains unchanged rather than elevated or lowered.","tokens_in":28973,"tokens_out":22536,"duration_ms":253480,"concrete_test":"Take T={0,1,2} and define an absolutely continuous path with an interior zero: let p_t(0)=0 for t<=0.5, p_t(0)=2(t-0.5) for t>0.5, and renormalize p_t(1),p_t(2) smoothly, so at t=0.5 we have p_t(0)=0 and \\dot p_t(0)=2. At that instant, Eq. 26 gives positive flux into state 0 from some positive-probability state, while the objective coefficient 1/(p_t(0)p_t(z)) is infinite. Check whether Eq. 26 satisfies the KKT stationarity condition (46a) with finite dual variables; if it does not, the claimed kinetic optimality for arbitrary p_t fails pointwise and must be replaced by an explicit regularization or a restriction to paths with p_t>0 on (0,1).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that Eq. 26 gives a kinetic-optimal flux for any prescribed discrete path rests on Proposition B.1, whose proof assumes p_t>0 pointwise. The extension to general p_t is justified by a 'safe flux' argument, but the kinetic energy objective (Eq. 19) with w_t(x,z)=1/p_t(x) assigns coefficient 1/(p_t(x)p_t(z)) to j_t(x,z). If at some time p_t(x)=0 while \\dot p_t(x)>0, any generating flux must carry positive mass into state x, so the objective is infinite at that instant; Eq. 26 is not the finite-energy KKT solution of Proposition B.1. This occurs, for example, at t=0 in the masked mixture path with a linear scheduler, where p_t(x_1)=0 but \\dot p_t(x_1)>0. The issue is measure-zero for the usual mixture and metric paths, so the velocity remains a valid generator, but the optimality claim for arbitrary support-changing paths is an ad hoc limiting construction rather than a proved theorem. This matters because the 'kinetic-optimal' label is used to justify preferring the Eq. 33 schedulers over masking: if Eq. 26 is only one convenient safe velocity, not the optimizer of a well-posed objective, then the mixture schedulers inherit a weaker justification.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a framework for discrete flow matching with arbitrary probability paths. It derives a closed-form kinetic-optimal velocity for any prescribed path (Eq. 26) and shows that optimizing the same objective over paths yields mixture paths with a source-dependent scheduler (Eq. 33). It also derives a tractable ELBO for mixture paths. Experiments on text, materials, and images show that the proposed paths and velocities often outperform the standard masking construction.","tokens_in":29315,"tokens_out":19811,"duration_ms":167117,"significance":"If the theoretical claims are fully established, the paper makes a significant contribution: it substantially enlarges the design space of discrete diffusion and flow models, provides a principled kinetic-energy criterion for selecting velocities and paths, and demonstrates practical gains across multiple modalities. The convex derivations for strictly positive paths are clean, the closed-form flux in Eq. (26) is elegant, and the tractable ELBO for mixture paths is a useful algorithmic contribution. The empirical validation is broad, spanning text, materials, and images, with several reproducible comparisons against strong baselines.","major_comments":[{"comment":"The statement that Eq. (26) is the kinetic-optimal flux for any prescribed probability path is proved only for strictly positive pt. The proof of Proposition B.1 explicitly assumes pt > 0 and uses the strict positivity of rho_t(x,z) = pt(z)/w_t(x,z). For paths with zero-probability states, the objective (19) with w_t(x,z)=1/p_t(x) assigns an infinite coefficient 1/(p_t(x)p_t(z)) to the flux j_t(x,z); if a generating flux must carry mass into a state with p_t(x)=0, as in the masked mixture path at t=0, the objective value is infinite and the optimization problem has no finite-energy solution. The 'safe flux' and limiting arguments given in the text do not establish that Eq. (26) is the minimizer of (19) in this regime. Since the kinetic-optimality label is used to justify the path and scheduler choices, the paper should either prove the zero-probability case by a rigorous approximation argument that preserves a well-defined sense of optimality, or explicitly restrict the optimality theorem to pt>0 and present the general case as a limiting construction.","section":"Section 4.1, Eq. (19)/(26); Appendix B, Prop. B.1"},{"comment":"The same positivity restriction affects the kinetic-optimal path derivation. Proposition B.2 assumes pt>0, and the energy equivalence used in its proof involves terms such as (d/dt sqrt(pt))^2 and intermediate expressions that become 0/0 at times when a state has zero probability, for example at t=0 for a mask source with p(x1)=0. Thus, the claim that mixture paths with the scheduler (33) are kinetic-optimal for the common mask-source case is unsupported by the proof as written. This claim is the main theoretical motivation for preferring these schedulers over masking, so the paper should clarify the exact status of (33) when p(x1)=0 and provide a rigorous justification, or explicitly weaken the claim to a heuristic or a limiting case.","section":"Section 4.2, Eq. (33); Appendix B, Prop. B.2"},{"comment":"The power-infinity velocity is defined by taking the limit alpha->infinity in Eq. (74), which yields the factor delta_{argmax_s p_t(s)}(x). This is not well-defined when the maximum of p_t is attained by more than one state, which occurs, for instance, for a uniform source at t=0. Without a tie-breaking rule, Eq. (77) is ambiguous, and the experiments in Figures 5-7 that use this velocity are not fully reproducible. The paper should specify how ties are broken or restrict the claim to paths with a unique maximizer.","section":"Appendix C.3, Eq. (77)"}],"minor_comments":[{"comment":"The notation \"∂tpt(x)\" in Eq. (26) is introduced without definition; please write \"∂_t p_t(x)\" and state that it denotes the time derivative.","section":"Section 4.1"},{"comment":"The sentence \"Indeed the above flux satisfy the Continuity Equation and the Rate Conditions as in Indeed the above flux satisfy the Continuity Equation and the Rate Conditions as in equation 17\" is duplicated and contains a subject-verb agreement error; it should be corrected.","section":"Appendix C.3"},{"comment":"The sentence \"we used linear and kinetic optimal schedulers with mask, p(x) = δm (x), and β0 ∈ {...} source distributions\" is confusing because of the comma after \"mask\"; please rephrase, for example \"with a mask source p(x)=δm(x) as well as β0 ∈ {...}\".","section":"Section 8.1"},{"comment":"The sentence \"Our model are on trained OpenWebText\" is ungrammatical; it should read \"Our models are trained on OpenWebText and FineWeb-Edu.\"","section":"Appendix E.1"},{"comment":"The text \"we sample t in [0, 1 − 1e−3]\" would be clearer with standard notation, e.g., \"we sample t in [0, 1−10^{-3}]\"; the same applies to similar expressions in the appendix.","section":"Appendix E.1"},{"comment":"The middle-panel labels \"u_t (p = 1)\" and \"u_t (p = ∞)\" are not explained in the caption; please define p or rename them to match the notation of Eqs. (74)-(77).","section":"Figure 2"},{"comment":"The metric-path model was trained for 600 epochs, while the masked and autoregressive baselines were trained for 300 epochs; this training budget discrepancy should be acknowledged in the comparison, as it may affect the reported FID differences.","section":"Section 8.4 and Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper presents an elegant and broadly validated framework, and the empirical results are strong. The main issue is the mismatch between the stated 'kinetic-optimal' claims and the proofs, which assume strictly positive paths; the zero-probability cases that are standard in masked discrete diffusion are not covered by the theorems. I recommend major revision so that the authors can either close this gap with a rigorous approximation argument or soften the optimality claims accordingly. The paper fits the journal's scope well."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should read this one. The headline result is a closed-form kinetic-optimal velocity for any discrete probability path (Eq. 26), which decouples the corruption process from the velocity; plus a source-dependent scheduler (Eq. 33) that beats masking on several benchmarks, and a tractable ELBO for mixture paths (Eq. 37). That is a real advance over the masked-only design space in prior work.\n\nWhat is actually new: the velocity formula is simple and does generate the prescribed path — they verify this directly, and it recovers previous velocities as special cases. The hypersphere-geodesic view of mixture paths is elegant, and the source-dependent scheduler appears novel. The ELBO is a useful contribution, giving a closed-form training objective for mixture paths that previously required doubly stochastic estimators.\n\nSoft spots: the \"kinetic-optimal\" label is only proven for strictly positive paths (Prop. B.1). For zero-probability states, the weight w_t = 1/p_t makes any flux into such a state have infinite energy, so the optimization problem is not well-posed at boundaries like t=0 in the masked path. The paper says the flux is safe and works for general p_t, but it does not prove it is the limit of the positive-path optimizers. That is a real gap, and it weakens the claim that Eq. 33 is the kinetic-optimal scheduler for the actual (support-changing) paths used in practice. The recommendation to prefer these schedulers over masking therefore rests partly on an ad hoc limiting argument — though the empirical results suggest the recommendation may still be right.\n\nAlso, the choice of objective weight is motivated by closed-form solvability and numerical stability, not by an independent principle. If that energy is not the right measure of path quality for discrete generation, the normative force of the schedulers fades, even though the velocity formula stands. Empirically: no code or checkpoints, no error bars, and the ImageNet metric-path hyperparameters were tuned visually on the same task. Several text gains over masking are within a point of perplexity, so \"outperform\" is fragile; the materials and image results are more striking but need reproduction.\n\nBottom line: the math core is coherent, the design space genuinely opens up, and the paper is honest about the safe-flux extension even if it underplays the gap. A serious referee should engage, primarily to pin down the zero-probability optimality claim and to request code, checkpoints, and error bars. I would cite it for the velocity formula and the ELBO, and I would bring it to reading group.","headline":"A genuinely useful generalization of discrete flow matching to arbitrary paths, with an optimality story that is solid for positive paths and heuristic for zero-probability boundaries; worth refereeing despite missing code and error bars.","tokens_in":29860,"tokens_out":6129,"would_cite":true,"duration_ms":58249,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Discrete flow matching need not be stuck with masking: any probability path can be generated by a closed-form kinetic-optimal velocity, and optimizing the same energy over paths selects mixture paths with a token-dependent scheduler.","keywords":["discrete flow matching","continuous-time Markov chains","kinetic-optimal paths","discrete probability paths","mixture paths","masked diffusion","evidence lower bound","source-dependent scheduler"],"falsifier":"Solve the convex kinetic problem (19) numerically on a small state space with $w_t=1/p_t(x)$ for a fixed positive path and compare with the closed-form flux (26); a mismatch would show the closed form is not the claimed minimizer. Reproducing the FineWeb-Edu text comparison, if a masked scheduler with any source distribution strictly dominates every kinetic-optimal scheduler on the ELBO (37), the superiority claim would fail.","tokens_in":28758,"feed_emoji":"🎲","tokens_out":8230,"duration_ms":70963,"temperature":0.7,"pith_summary":"Discrete flow matching is a way to generate token sequences by simulating a continuous-time Markov chain whose time marginals follow a prescribed probability path, and so far the field mostly uses a masked corruption process. This paper claims that any probability path the user writes down can be generated by a closed-form kinetic-optimal velocity, which decouples the choice of corruption process from the choice of velocity. It further claims that optimizing the same kinetic energy over probability paths selects mixture paths with a token-dependent scheduler, and that a new tractable evidence lower bound (ELBO) trains these paths. If these claims hold, practitioners can design corruption processes from domain knowledge, and kinetic-optimal mixture paths outperform masking on text, materials, and image generation.","feed_headline":"One formula puts any discrete noise schedule into flow matching","feed_subtitle":"Kinetic-optimal velocities let users pick any corruption process and beat masking on text and images.","key_machinery":"The central object is the forward flux $j_t(x,z)=u_t(x,z)p_t(z)$ and the discrete kinetic energy $\\sum_{x\\ne z} w_t(x,z)j_t(x,z)^2/p_t(z)$. With a symmetric weight the constrained minimization relaxes to a linear Laplacian system (21); the choice $w_t(x,z)=1/p_t(x)$ makes that system solvable in closed form, giving the flux (26) and its velocity. The same weight turns path optimization into a geodesic on the sphere of square-root probabilities, yielding the source-dependent scheduler (33). The tractable ELBO (37) completes the machinery by providing the training objective for mixture paths.","core_discovery":"The paper establishes a complete design space for discrete flow matching. For a fixed strictly positive probability path $p_t$, the flux $j_t(x,z)=u_t(x,z)p_t(z)$ that minimizes the symmetric kinetic energy with weight $w_t(x,z)=1/p_t(x)$ is given in closed form by $j^*_t(x,z)=[p_t(z)\\dot p_t(x)-\\dot p_t(z)p_t(x)]_+$, and converting this flux to a velocity yields a safe generator for $p_t$ (Eq. 26). When the probability path itself is optimized under the same energy, the problem becomes geodesic motion on the sphere of square-root probabilities; for conditional paths it recovers the mixture path with the source-dependent scheduler $\\kappa_t(x_1)=1-\\sin^2((1-t)\\Omega)/\\sin^2\\Omega$, where $\\Omega=\\arccos\\sqrt{p(x_1)}$. The paper also derives an evidence lower bound for mixture paths, Eq. 37, which is tractable and contains the masked ELBO as a special case. Across text, crystal, and image benchmarks, these kinetic-optimal paths and metric-induced paths match or beat the masked construction, with the largest gains in low-budget sampling and in permutation-invariant crystal generation.","pith_inferences":["Editorial inference: the same machinery could be applied to other discrete spaces with natural distances, such as audio tokens or molecular graphs, by plugging a metric into the metric-induced path (27).","Editorial inference: the proof of optimality depends on the chosen weight $1/p_t(x)$; a data-driven or task-driven weight would likely produce different schedulers while leaving the fixed-path velocity formula intact.","Editorial inference: the paper's split into a probability-advancing flux and a probability-preserving corrector suggests a general sampling strategy, kinetic-optimal velocity at coarse steps plus symmetric corrector flux at fine steps, which could be tested independently on other discrete generators."],"forward_implications":["Any user-specified discrete probability path, whether mask, uniform, metric-induced, or bespoke, now has an explicit safe generating velocity, so corruption-process design is decoupled from velocity design.","The kinetic-optimal scheduler is source-dependent, so non-mask source distributions become competitive with masking for text, not just equal to it.","The tractable ELBO (37) gives mixture-path models a likelihood bound for training and evaluation, and reduces to the masked ELBO as a special case.","Metric-induced paths improve quality at low numbers of sampling steps and beat the masked baseline on CIFAR-10 and face-blurred ImageNet-256.","Permutation-invariant discrete flow matching with kinetic-optimal schedulers reaches state-of-the-art stability rates in inorganic crystal generation."],"supporting_citations":[{"why":"Supplies the discrete kinetic energy and optimal-transport objective (Eq. 19) that the paper optimizes.","marker":"Peyré et al. (2019)"},{"why":"Defines discrete flow matching with mixture paths and the velocity recovered by Eq. (26), and provides the main mask baseline for comparisons.","marker":"Gat et al. (2024)"},{"why":"Gives the continuous-time Markov-chain flow framework and the marginal-velocity construction (Eq. 9) that this paper generalizes.","marker":"Campbell et al. (2024)"},{"why":"Motivates kinetic-optimal probability paths in the continuous setting, which the paper transfers to discrete state spaces.","marker":"Shaul et al. (2023)"},{"why":"Provides the masked mixture ELBO that the paper recovers in Appendix D.1, and a key text-generation baseline.","marker":"Shi et al. (2024)"},{"why":"Shows masked diffusion behaves like any-order autoregression and cautions on low-precision sampling, motivating non-mask paths.","marker":"Zheng et al. (2024)"},{"why":"The continuous flow-matching framework whose design-space decoupling is mirrored here.","marker":"Lipman et al. (2022)"}],"fun_headline_variants":["Closed-form velocity for any discrete flow path","Optimal discrete path formula beats masking","One kinetic-optimal formula for discrete flows","Any corruption process now has an optimal velocity","Mixture paths: kinetic-optimal for discrete flows"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The optimality results stand on a particular symmetric kinetic energy with weight $1/p_t(x)$, chosen because it yields closed forms and numerical safety rather than because it is the right measure of path quality.","fun_headline_variants_meta":{"raw":{"variants":["Closed-form velocity for any discrete flow path","Optimal discrete path formula beats masking","One kinetic-optimal formula for discrete flows","Any corruption process now has an optimal velocity","Mixture paths: kinetic-optimal for discrete flows"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000211,"raw_usage":{"total_tokens":1438,"prompt_tokens":994,"completion_tokens":444,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":377}},"tokens_in":610,"tokens_out":444,"duration_ms":4804,"temperature":1.0,"reasoning_tokens":377,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:20:49.649606+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve the convex kinetic problem (19) numerically on a small state space with $w_t=1/p_t(x)$ for a fixed positive path and compare with the closed-form flux (26); a mismatch would show the closed form is not the claimed minimizer. Reproducing the FineWeb-Edu text comparison, if a masked scheduler with any source distribution strictly dominates every kinetic-optimal scheduler on the ELBO (37), the superiority claim would fail.","supporting_citations":[{"cited_title":"(2024): (i) We replace the first layer with an embedding table of size 256 × 96, and we stack the channel features such that the input to the U-Net is of shape 288 × 32 × 32","cited_arxiv_id":null,"evidence_quote":"Defines discrete flow matching with mixture paths and the velocity recovered by Eq. (26), and provides the main mask baseline for comparisons."}],"review_version":1}