{"id":"102e75e4-2045-4a3b-8c88-db50b224d5c1","arxiv_id":"2607.24522","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Trajectory-derived, temporally weighted velocity matching from shared student states outperforms KL-based on-policy distillation for multi-reference flow model post-training.","lead":"FlowCTS trains flow image models by matching student and teacher trajectories from the same student state, not by KL on SDE kernels. It beats KL-based on-policy distillation on GenEval, OCR, and PickScore and exposes a late-timestep bias in the KL baseline.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"At K=1 the entire CTS derivation collapses to plain unweighted velocity MSE at a student-visited state, so the paper's causal story (trajectory-derived bound/weights beat KL OPD) is never isolated from the much simpler explanation already offered by concurrent DanceOPD.","rationale":"The reader's weakest_assumption identifies the same soft spot: empirical isolation of the weights/bound vs plain MSE is incomplete, and the K=1 objective coincides with DanceOPD. My pass confirms this is the genuinely load-bearing point and sharpens it: the issue is not just that the bound might be loose (the reader's framing), but that at the K where all headline numbers are produced, the bound and weights are literally inert (α_0=1), so the paper's mechanistic attribution is supported only by the correlational Fig. 4 and never by a controlled weighted-vs-unweighted ablation. This does not, however, undermine the controlled empirical claim that unweighted on-policy velocity matching beats transition-kernel KL OPD under matched training — Table 1 and Fig. 3 are legitimate within-pipeline evidence for that narrower statement, and nothing in the derivation is circular or contradictory. The appropriate verdict remains CONDITIONAL: accept-shaped if the authors (i) add the weighted-vs-unweighted isolation experiment, (ii) add seeds/error bars, and (iii) reframe the contribution as the temporal weighting analysis plus the K-family study rather than the bound itself, given the K>1 results show no consistent benefit. Reject would be too strong; accept would let the unsupported causal framing stand.","tokens_in":16695,"tokens_out":1767,"duration_ms":43859,"concrete_test":"Run two matched variants in the identical OPD pipeline: (a) uniform-weight velocity MSE at student-visited states (=FlowCTS K=1), and (b) velocity MSE weighted by the same w_KL(t) profile as Vanilla OPD but computed directly on the ODE states (no SDE-kernel machinery — Eq. 11 shows these are equivalent up to the weight). If (a) ≈ (b), the temporal-misallocation story in §4.4 fails and the gain must come from some other confound (objective scale, EMA, gradient norm interaction); if (a) > (b), the diagnosis is confirmed and the paper should say the contribution is the unweighting, not the trajectory bound. Separately, rerun the Table 1 K=1 comparison with 3 seeds and report mean±std; if the GenEval 0.90→0.93 gap is within one std, the headline empirical lift itself needs hedging.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim attributes the gains over Vanilla OPD to the trajectory-derived objective: the VUB (Eq. 8), its Euler discretization with weights α_i = [K(K+1)−i(i+1)]/2 (Eq. 10/33), and the temporal-misallocation diagnosis of Fig. 4. But the headline result is at K=1, where α_0=1 and the objective is exactly mean-squared velocity matching between student and reference at one on-policy state — the same objective DanceOPD (Zhou et al., 2026) motivates as local field regression. The Cauchy–Schwarz bound and the α-weight profile do no work at K=1. The multi-step variants that would actually exercise the trajectory machinery (K=2,3) give inconsistent or negative deltas (PickScore 23.06→22.60 at K=3; Table 1), so the paper provides no evidence that the bound/weights contribute anything measurable. What the K=1-vs-Vanilla-OPD comparison actually isolates is one change: removing the SDE-induced weight w_KL(t) from velocity regression (per Eq. 11, the KL objective *is* weighted velocity regression). Fig. 4 supports the diagnosis that w_KL misallocates supervision late, but the paper never tests the clean counterfactual — weighted vs. unweighted velocity regression at identical states — so the 'why' rests on a correlational temporal plot, and the framing 'trajectory supervision is why we win' is not supported by any ablation distinguishing it from plain uniform velocity MSE. Secondarily, the absolute lifts (GenEval +0.03, OCR +0.02, PickScore +0.31) are reported with a single seed (seed 42, App. D) and no error bars; GenEval is known to fluctuate by ~0.01–0.02 across seeds at this scale, so even the empirical premise of the attribution is on the edge of noise.","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper proposes FlowCTS, an on-policy distillation objective for flow models that supervises the discrepancy between student and reference ODE trajectories launched from a shared student-visited state. Starting from the integral identity between trajectory and velocity fields, the authors derive a Cauchy–Schwarz velocity upper bound (VUB, Eq. 8) with a trajectory-derived weight w(r), and discretize it over K Euler steps into a weighted velocity-matching objective with coefficients α_i = [K(K+1)−i(i+1)]/2 (Eq. 10, App. B). In a multi-reference setup on SD3.5-Medium with task-specific Flow-GRPO experts as references, K=1 FlowCTS-OPD reportedly improves over a transition-kernel KL baseline (Vanilla OPD, following Flow-OPD) on GenEval (0.90→0.93), OCR (0.90→0.92), and PickScore (22.75→23.06), with faster convergence (Fig. 3) and better OOD compositional and preference metrics (Tables 2–3). The authors diagnose the KL baseline's weakness as a temporal supervision mismatch: the SDE-induced weight w_KL(t) concentrates supervision late in denoising while student–reference velocity gaps peak early (Fig. 4). An off-policy SFT variant and a study of supervision horizon K are also reported.","tokens_in":17078,"tokens_out":2898,"duration_ms":98057,"significance":"If the results hold, the paper makes a useful contribution to an active area: a controlled multi-reference OPD comparison in which prompt mixture, rollouts, reference routing, and optimization are held fixed between objectives (§4.1, App. C.1), a clean derivation of a trajectory-discrepancy upper bound with trajectory-derived weights (Apps. A–B), a training-budget control for the K=2 comparison (App. E), and a temporal-misallocation analysis of KL-based Flow-OPD (§4.4, Fig. 4) that is consistent with, and adds evidence to, the TempFlow-GRPO timing story. The off-policy extension and OOD evaluations on T2I-CompBench and DrawBench broaden the evidence. However, the significance is currently capped by two factors: the objective at the headline operating point (K=1) coincides with the concurrent DanceOPD field-regression objective, so the distinctive \"trajectory\" machinery is not what delivers the reported gains; and the multi-step variants that do exercise the machinery give non-monotonic or negative effects (PickScore 22.60 at K=3). The honest, still-valuable version of this paper is \"unweighted velocity regression at student-visited states beats KL-weighted OPD, and here is a bound/","major_comments":[{"comment":"The headline comparison is at K=1, where α_0=1 and Eq. (10) is exactly unweighted mean-squared velocity matching between student and reference at a single student-visited state — by the authors' own Eq. (11), the only difference from Vanilla OPD is the removal of the SDE-induced weight w_KL(t). The VUB derivation (Eq. 8, App. A) and the α-weight profile (Eq. 33, App. B) therefore do no work in the headline experiment, and the K=2,3 rows that would exercise the trajectory machinery show inconsistent or negative effects (PickScore 23.06→22.60 from K=1 to K=3; OCR flat). The abstract and §1 attribute the gains to the 'trajectory-derived velocity-matching objective,' but no experiment separates that attribution from the simpler hypothesis 'uniform velocity regression beats KL-weighted velocity regression.' The clean counterfactual — KL-weighted vs. unweighted velocity regression at identical","section":"§3.3, Eq. (10); §4.2, Table 1; §4.4, Eq. (11)"},{"comment":"All headline deltas (GenEval 0.90→0.93, OCR 0.90→0.92, PickScore 22.75→23.06) are reported from what appears to be a single training run per configuration, with no seed variance, confidence intervals, or significance testing. GenEval and OCR are known to have non-trivial evaluation variance, and a 0.02–0.03 gap is within plausible run-to-run fluctuation for LoRA fine-tuning. The same concern applies to Table 3 (e.g., ImageReward 1.21→1.27) and is acute in Table 2, where Vanilla OPD actually wins Color (0.8350 vs. 0.8295), Texture, and Non-Spatial, undercutting the claim of consistent OOD improvement. At minimum, the main Table 1 comparison and Fig. 3 curves should be replicated over ≥3 seeds (or the evaluation variance quantified by repeated decoding of fixed checkpoints), and claims of 'consistent' improvement should be restricted to metrics where the gap exceeds measured noise.","section":"§4.2, Table 1; §4.5, Tables 2–3"},{"comment":"The VUB is obtained by one application of Cauchy–Schwarz followed by an integral reordering (App. A, Eq. 17); its tightness depends on the velocity mismatch being roughly aligned along the segment, which fails as the student and reference paths diverge. The paper never quantifies how loose L_VUB is relative to L_CTS, nor does it establish that minimizing the bound improves the true trajectory objective. This matters most exactly where the bound is stressed — K>1 — and the non-monotonic K results in Table 1 are consistent with the bound becoming uninformative as paths diverge. A diagnostic (e.g., reporting the empirical ratio L_CTS/L_VUB or the trajectory discrepancy at the end of training for K=1,2,3) would either substantiate or honestly bound the mechanism claim. Relatedly, the discrete derivation (App. B) applies Cauchy–Schwarz per step and then discards the h^3 factor; a brief note o","section":"§3.2, Eq. (8); Appendix A, Eq. (17)"}],"minor_comments":[{"comment":"Abstract/§1: 'Beyond on-policy setting,FlowCTS' is missing a space; similar spacing artifacts ('from0.90to0.93', 'Vanilla SFT ,') appear throughout, likely from PDF extraction but worth fixing in source.","section":"Abstract"},{"comment":"The off-policy gains over Vanilla SFT are marginal and partly negative: at K=2 PickScore drops 21.68→21.66 and GenEval moves only 0.71→0.72; the 'consistently outperforms vanilla SFT' claim in the abstract rests almost entirely on OCR (+0.03–0.05). The abstract should reflect this metric-selective pattern.","section":"§4.2, Table 1 (Part II)"},{"comment":"Table 4 uses a 'reduced training configuration' and its K=2 numbers (GenEval 0.93, PickScore 22.80) do not match the Table 1 K=2 row (0.94, 22.96). Please state explicitly what differs between the two configurations and caution that Table 4 is not comparable to Table 1.","section":"Appendix E, Table 4"},{"comment":"The phrase 'a new scaling law dimension' (§1, contributions) overstates what is shown; three values of K with non-monotonic outcomes is a hyperparameter study, not a scaling law. Suggest 'a new design dimension' or similar.","section":"§1, contributions list"},{"comment":"Fig. 4 would be more convincing with the w_KL(t) weight itself plotted alongside panels (a) and (b), so the reader can see directly that the concentration of the KL loss is weight-driven rather than loss-driven. Also specify over how many states/prompts the curves are averaged.","section":"§4.4, Fig. 4"},{"comment":"§4.1/Appendix C: the timestep fraction 0.99 yielding 'nine denoising transitions' and the 1:3:1 prompt-cycle ratio are load-bearing hyperparameters with no sensitivity analysis; a brief ablation or at least a justification for these choices would help reproducibility. Code/checkpoint release plans are not stated.","section":"§4.1; Appendix C.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript sits in a crowded concurrent space (Flow-OPD, DiffusionOPD, DanceOPD, Uni-OPD, all 2026 preprints). The authors disclose the DanceOPD overlap, which is appropriate, but reviewers should note that at K=1 — the setting of the headline result — the objective is functionally identical to DanceOPD's field regression, so the paper's distinct contribution rests almost entirely on the K>1 formulation and the temporal-misallocation analysis, the former of which currently shows weak empirical value. The citation pattern is otherwise normal for the area. Fit with the journal seems fine assuming the venue accepts flow-model post-training work."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The thing worth knowing: under a matched multi-reference OPD setup on SD3.5-M, plain velocity matching at student states beats transition-kernel KL OPD (GenEval 0.90→0.93, OCR 0.90→0.92, PickScore 22.75→23.06) with faster curves, and Fig. 4 gives a clean reason—KL mass piles up late while student–reference velocity gaps peak early. That temporal diagnosis is the sharpest part of the paper and lines up with TempFlow-GRPO intuition.\n\nWhat is actually new is less the K=1 objective than the package around it: the path-integral VUB and discrete α weights, the K-family framing, the off-policy SFT extension (OCR gains), and especially the KL-vs-velocity-gap plot. Math in Apps. A–B is standard Cauchy–Schwarz / sum reordering and internally consistent. The Vanilla OPD comparison holds prompts, rollouts, routing, and optimizers fixed, so the objective-level claim is fair as far as it goes. Multi-ref balance vs single-task GRPO experts and Mix is a useful empirical point for people doing flow post-training.\n\nSoft spot, in proportion: the stress-test is right. At K=1, α0=1 and CTS is exactly unweighted velocity MSE—the same local field regression DanceOPD already motivates. The bound and trajectory weights do no work in the headline result. K=2/3, which would exercise the machinery, are mixed (PickScore drops at K=3). So “trajectory supervision is why we win” is not isolated from “drop the SDE-induced w_KL(t).” They never run weighted vs unweighted velocity MSE at identical states. Lifts are single-seed, no error bars, absolute deltas small. Framing oversells the continuous-trajectory derivation relative to what Table 1 actually isolates.\n\nWho it’s for: people post-training flow/diffusion generators with OPD or multi-expert distillation. Not a theory paper; a methods + diagnosis note. I’d bring it to reading group only if we’re deep in flow alignment that week. It deserves peer review—solid enough controlled experiment and a real temporal insight—but referees should force the weight ablation, multi-seed stats, and a clearer split from concurrent velocity-matching OPD. Engage if you care about OPD objectives for flows; skim Fig. 4 and Table 1 if you only need the takeaway.","headline":"Real win is unweighted on-policy velocity match vs SDE-KL weighting; the trajectory-bound story is mostly unused at the headline K=1 result.","tokens_in":17937,"tokens_out":616,"would_cite":true,"duration_ms":18924,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Matching student and reference trajectories from shared on-policy states beats KL distillation for flow models.","keywords":["flow matching","on-policy distillation","trajectory supervision","velocity matching","text-to-image","post-training","SDE transition kernels"],"falsifier":"A controlled ablation that replaces the trajectory-derived weights with uniform or late-heavy weights (or plain unweighted velocity MSE at the same on-policy states) and checks whether the reported gains over KL-OPD and the early-timestep advantage disappear.","tokens_in":17486,"feed_emoji":"🌊","tokens_out":896,"duration_ms":20520,"temperature":0.7,"pith_summary":"Flow models generate images by integrating a continuous velocity field from noise to data, so they lack the natural next-token distributions that make on-policy distillation work for language models. This paper argues that the right dense supervision is continuous trajectory matching: from a state the student actually visits, force the student’s subsequent path to stay close to a reference expert’s path. From the integral link between paths and velocities they derive a weighted velocity-matching upper bound, then discretize it into practical losses controlled by how many future steps K are supervised. In a multi-expert on-policy setup the single-state version already beats transition-kernel KL distillation with faster convergence and higher GenEval, OCR, and PickScore, and also beats a mixed-reward RL baseline; the same idea improves off-policy fine-tuning, especially OCR. The authors locate the gain in better early-denoising supervision: KL weights from auxiliary SDEs pile loss near the end of sampling, exactly where student–reference velocity gaps are smallest.","feed_headline":"Trajectory matching beats KL for on-policy flow distillation","feed_subtitle":"Shared-state path supervision lifts GenEval, OCR, and PickScore and fixes late-heavy KL weights","key_machinery":"Flow Continuous Trajectory Supervision (FlowCTS): the integrated squared gap between student and reference trajectories started from the same state, bounded by a trajectory-derived weighted velocity MSE (the velocity upper bound) and discretized into a K-step weighted velocity-matching loss.","core_discovery":"On-policy distillation for flow models is more effective when formulated as continuous trajectory supervision from shared student-visited states than as reverse-KL on auxiliary SDE transition kernels. The trajectory discrepancy admits a temporally weighted velocity-matching upper bound whose Euler discretization yields a family of objectives parameterized by supervision horizon K; K=1 already outperforms vanilla KL-OPD with faster convergence on GenEval, OCR, and PickScore, while longer K trades richer path information against harder optimization.","pith_inferences":["If early-path supervision is the real driver, any OPD or RL method whose effective weight peaks late (SDE kernels, terminal rewards) is systematically misallocated for flow and diffusion models.","Concurrent direct velocity matching at student states may capture much of the K=1 gain; the distinctive claim to pressure-test is whether the derived multi-step weights and path rollouts add value beyond local field regression.","Treating short trajectory segments as the supervision atom could transfer to video or other long-horizon continuous generators where exposure bias compounds over many steps."],"forward_implications":["Dense OPD for continuous generators should target early denoising stages, where student–reference velocity gaps and structural decisions are largest.","Supervision horizon K becomes an explicit scaling knob for flow post-training, balancing path information against optimization difficulty.","The same trajectory objective applies off-policy: multi-step FlowCTS-SFT can beat single-state velocity SFT, especially on text rendering.","Multi-reference capability consolidation can be done by routing student states to task experts under one trajectory loss instead of mixed sparse rewards."],"fun_headline_variants":["Shared-state trajectory matching beats KL in on-policy flow distillation","FlowCTS: continuous path supervision outperforms KL-OPD on flow models","Velocity-matching from shared states lifts GenEval OCR and PickScore","Temporal trajectory bound fixes late-heavy weights in flow OPD","K=1 FlowCTS-OPD converges faster than vanilla KL on-policy distillation"],"cache_read_input_tokens":128,"weakest_assumption_plain":"Optimizing the Cauchy–Schwarz velocity upper bound and its short Euler discretization is assumed to be a faithful enough stand-in for true path matching, rather than gains coming mainly from plain on-policy velocity regression without that bound.","fun_headline_variants_meta":{"raw":{"variants":["Shared-state trajectory matching beats KL in on-policy flow distillation","FlowCTS: continuous path supervision outperforms KL-OPD on flow models","Velocity-matching from shared states lifts GenEval OCR and PickScore","Temporal trajectory bound fixes late-heavy weights in flow OPD","K=1 FlowCTS-OPD converges faster than vanilla KL on-policy distillation"]},"model":"grok-4.5","effort":"low","cost_usd":0.002181,"raw_usage":{"total_tokens":960,"prompt_tokens":786,"num_sources_used":0,"completion_tokens":78,"cost_in_usd_ticks":21808000,"prompt_tokens_details":{"text_tokens":786,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":96,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":786,"tokens_out":78,"duration_ms":3251,"temperature":1.0,"reasoning_tokens":96,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T12:27:57.347459+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"A controlled ablation that replaces the trajectory-derived weights with uniform or late-heavy weights (or plain unweighted velocity MSE at the same on-policy states) and checks whether the reported gains over KL-OPD and the early-timestep advantage disappear.","supporting_citations":[],"review_version":1}