{"id":"548eae1e-61e3-41de-9460-3d88ac2f1247","arxiv_id":"2607.16361","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Integration error in generative ODEs is injected where trajectory variation is high and then transported across regions; per-region error is partly predicted by Flow Complexity and reconstructed by propagated signed truncation residuals.","lead":"Where does coarse-solver error enter a generated image, and how does it travel? This paper shows a per-region 'Flow Complexity' score read from the model's own velocity field predicts where the error lands, and that signed truncation residuals propagated through the model's linearized dynamics reconstruct the endpoint error. Because few-step sampling is how these models are deployed, the map could guide adaptive solvers and training.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reconstruction's first-order closure is not checked against actual cheap-trajectory divergence; the 0.81–0.87 cosine may reflect reference-path linearization rather than the real transport of integration error.","rationale":"The reader identified the teacher-forced residual definition as the weakest assumption. I agree that the reconstruction's validity is the most load-bearing point, but I would sharpen the mechanism: the evaluation point of τ_k is standard in numerical analysis; what is genuinely unverified is whether the linearized propagation around the reference trajectory closes on the actual nonlinear cheap-solve error. The paper's NFE sweep and linear-response checks are reassuring but do not report trajectory divergence or the nonlinear remainder. The central claim—that per-region error is local injection redistributed by transport—depends on the reconstruction being a faithful first-order account of the actual endpoint error. If the reconstruction is instead measuring a reference-path quantity, the diagonal-vs-transported contrast and the null results lose their bearing on real sampling. The missing divergence check is addressable and does not in itself demonstrate an error, so the reader's CONDITIONAL verdict remains appropriate. No code/data are released, which further supports keeping the verdict conditional rather than accept.","tokens_in":21998,"tokens_out":13560,"duration_ms":163531,"concrete_test":"On the two reconstruction models, report the per-step relative divergence d_k = ||x_k^cheap − x_ref_k|| / (1 + ||x_ref_k||). Then recompute the reconstruction two ways: (i) evaluate residuals at the actual cheap states, using the same near-exact solver to define the local defect; (ii) propagate the variational equation with J evaluated along the cheap trajectory instead of the reference trajectory. If the endpoint cosine and per-region ρ remain at the reported levels (≥0.8 and ≥0.7), the reference-path concern is resolved. If they fall toward the no-propagation baseline (0.71/0.77), the reported closure is an artifact of linearizing around the reference path rather than the actual transport of integration error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central transport claim rests on the reconstruction in Appendix A.2: the actual endpoint error e is approximated as Σ Φ(T,t_{k+1}) τ_k, where τ_k is the teacher-forced Euler defect evaluated at the reference state (Eq. 4) and Φ is estimated by finite differences of the near-exact solve. This is a standard local-truncation expansion, but its quantitative validity depends on the cheap Euler trajectory staying close enough to the reference trajectory that the linearized propagation captures the actual nonlinear error. The paper never reports the per-step divergence between the cheap and reference trajectories, nor the magnitude of the neglected nonlinear remainder. For few-step generative ODEs (NFE 4–6), the cheap trajectory can diverge substantially from the reference; if that divergence is large, the 0.81–0.87 endpoint cosine and the diagonal-vs-transported contrast (ρ≈0.33–0.42 vs partial ρ≈0.66–0.69) could be dominated by the reference-path linearization rather than by the actual error-injection and transport process. The conclusion that 'a region's error owes more to what arrives from elsewhere' is exactly the load-bearing inference; if the reconstruction is a reference-path artifact, that inference loses its quantitative support.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies where integration error is injected and how it spreads when a trained generative ODE is solved with very few steps. It defines a per-region diagnostic, Flow Complexity (FC), as the total variation of the model's velocity/prediction field along the trajectory, and shows that FC predicts the per-region reference-solver discrepancy (cheap solve vs. near-exact solve) with within-image Spearman rho 0.57–0.70 across five 256px models. It then uses perturbation experiments and a signed-residual reconstruction (e ≈ Σ Φ τ_k) to argue that endpoint error is mostly transported from other regions rather than locally injected, with off-diagonal contributions carrying more association than a region's own injection. A training penalty on FC lowers reference-solver discrepancy and improves few-step FID at matched optimization steps on the authors' own models, though not at matched training compute on ImageNet. The paper is notably explicit about its limitations and scope.","tokens_in":22306,"tokens_out":5886,"duration_ms":96908,"significance":"If the central transport claim holds, the paper is a genuinely new spatial account of numerical error in generative ODEs: it separates where error is injected from where it lands, shows that the source can be read from the model itself, and demonstrates that the source structure can be modified by training. The paper's strengths include convergence checks (rank correlation ≥0.986 against finer references), image-level bootstrap CIs, protocol-frozen nulls, matched-budget selector comparisons, and explicit acknowledgment of single-run and compute-matched caveats. The empirical apparatus is unusually careful, and the FC landscape result alone is a useful contribution. However, the load-bearing reconstruction and the diagonal-vs-transported contrast rest on a reference-path linearization whose validity is not demonstrated against actual cheap-trajectory divergence.","major_comments":[{"comment":"The reconstruction defines the signed per-step truncation τ_k as the teacher-forced Euler defect evaluated at the reference state x_ref_k, and propagates it via linearized dynamics evaluated at x_ref_{k+1}. The target is the actual error of the cheap Euler solve, but the manuscript never reports how far the cheap trajectory diverges from the reference trajectory, nor the magnitude of the neglected nonlinear remainder. If the cheap trajectory diverges substantially over NFE 4–6, τ_k can differ from the defect actually injected during real sampling, and the 0.81–0.87 endpoint cosines and the diagonal-vs-off-diagonal contrast could be properties of the reference path rather than of the true error-injection process. The manuscript itself lists \"evaluation along the reference rather than the perturbed trajectory\" as a residual source (A.2), but gives no quantitative estimate. Please report pe","section":"A.2, Eq. (4)–(5)"},{"comment":"The headline contrast — own injection ρ≈0.33–0.42, transported-in partial ρ≈0.66–0.69 — is computed entirely from the same reference-path decomposition described above. The source-count, sign-randomization, and destination-permutation nulls are well designed, but they all operate on the reference-path objects. If those objects are not faithful to the actual cheap trajectory, the contrast may be a linearization artifact rather than evidence of real transport. The paper should provide a check in which the actual cheap trajectory is used, or at least show that the reconstruction's conclusions are stable when the propagation path is perturbed by the observed level of cheap-reference divergence. Absent that, the mechanistic interpretation of the partial correlations is under-supported.","section":"§6 / Fig. 6B"},{"comment":"The paper says the decomposition is \"only worth the name if it holds quantitatively,\" and uses \"closes\" language, yet the median relative error at the operational NFE=6 budget is 0.65 on CelebA and 0.50 on ImageNet. The cosines of 0.81–0.87 show that the first-order reconstruction captures direction and regional ordering, but roughly half to two-thirds of the error vector remains unexplained. Please state this directly in §6 as well as in A.2, and either decompose the residual (nonlinear defect interactions, finite-difference error, reference-solver error, reference-path mismatch) or soften the \"closes\" wording. This is relevant because the off-diagonal partial correlation is presented as the quantitative foundation of the main conclusion.","section":"A.2 (\"first-order closure\")"}],"minor_comments":[{"comment":"The ImageNet reconstruction sample size is inconsistent: §6 and Fig. 6B say n=24, while A.2 first reports n=32 before later switching to n=24. Please harmonize the reported n and update the CIs if needed.","section":"§6 / Fig. 6B / A.2"},{"comment":"The legend reads \"CelebA DiT (flow, latent)\" but DiT is described elsewhere as a diffusion model and SiT as the flow model; this is likely a typo for the authors' FM-DiT model. Please correct the legend to avoid confusion.","section":"Fig. 7A"},{"comment":"The reproducibility section states that code and configs \"will be released upon publication.\" Given the number of committed artifacts (results/sit/, results/celebahq/, nulls, baselines), a public repository link or archive with the analysis code would materially strengthen the paper.","section":"B.11"}],"recommendation":"major_revision","confidential_remarks":"The paper is unusually honest about its limits, and I do not see grounds for rejection. The central reconstruction, however, currently rests on an unquantified reference-path assumption; the authors can address this with a divergence report and a robustness check. The ImageNet compute-matched result is already framed correctly, so it does not block acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper deserves a serious referee. It gives the first spatially resolved, empirically closed account I know of for where few-step integration error is injected in generative ODEs and how it reaches the endpoint. The core claims are: (1) a per-region Flow Complexity statistic — total variation of the model's velocity/prediction along a fine reference trajectory — predicts the per-region cheap-vs-reference discrepancy with within-image Spearman 0.57–0.70 on five 256px models; and (2) a perturbation protocol shows a local disturbance spreads widely (under 10% of summed response stays at the source near the noise end), with a signed-residual reconstruction capturing endpoint error direction at cosine 0.81–0.87. Those numbers are backed by unusually good controls: convergence-checked references, image-level bootstrap CIs, protocol-frozen nulls, and honest reporting of where the method falls short.\n\nWhat's actually new is the spatial resolution. Classical numerical analysis already says global error is local truncation carried by the state-transition operator; the contribution here is turning that abstraction into an empirically validated per-region diagnostic and a transport measurement. The authors are honest about limits: they say FC is not a new dynamical quantity, they demote the training intervention to a construct-validity check, and they report the counter-result that FID-optimal allocation is the reverse of error-optimal allocation. That is good faith.\n\nSoft spots, in proportion. First, the reconstruction in Appendix A.2 evaluates the per-step truncation residual at reference states, not at the states the cheap solve actually visits. The paper never reports how far the cheap trajectory deviates from the reference. For NFE 4–6 that deviation can be substantial, so the 0.81–0.87 cosine might be partly a reference-path linearization artifact rather than a faithful account of the actual injection-and-transport process. The perturbation experiment independently establishes transport, so the qualitative conclusion likely survives, but the diagonal-vs-transported contrast (the load-bearing part) is less airtight than it looks. This is fixable — report per-step divergence, or recompute residuals at the cheap states — and I'd want it before fully trusting the quantitative reconstruction. Second, no code or data is released; for a paper this empirical that is a real limitation. Third, the 256px intervention arms are single-run, a caveat the authors acknowledge and partially mitigate with a three-seed controlled replication.\n\nOverall, the paper is careful, honest, and the central empirical claims are supported by the reported evidence. The missing divergence measure is the main soft spot, but it is a gap, not a demonstrated error. I would send this to peer review: it deserves a thorough referee, and the discussion it will generate is worth the time.","headline":"A careful, well-scoped empirical study: per-region integration error is predicted by Flow Complexity and is transported nonlocally, but the reconstruction's reference-path assumption needs a divergence check before the strongest claims fully land.","tokens_in":22745,"tokens_out":3606,"would_cite":true,"duration_ms":40502,"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":"Few-step integration error in generative ODEs is injected where the model's velocity or prediction field varies most, then transported far from that source by the learned dynamics; a signed propagation of per-step truncation residuals recon","keywords":["generative ODEs","flow matching","diffusion models","integration error","spatial error transport","Flow Complexity","error prediction","ODE solvers"],"falsifier":"Measure the divergence between an NFE-6 Euler path and the fine reference path on the paper's own models, and recompute the reconstruction with residuals evaluated along the cheap path instead of the reference path: if the endpoint cosine (0.81-0.87) and per-region rho (0.80-0.82) drop to the no-propagation level, the transport account describes reference-path error rather than the error experienced by the actual solve.","tokens_in":21876,"feed_emoji":"🖼️","tokens_out":7534,"duration_ms":68940,"temperature":0.7,"pith_summary":"This paper tries to establish where integration error enters when a trained flow or diffusion model is run with few solver steps, and how that error reaches the final sample. It claims the error is not noise: each image has a stable, learned spatial difficulty landscape, concentrated in a few regions and reproducible across budgets and retraining, and this landscape can be read from the model's own velocity or prediction field. The reading is only partial because error is injected locally but then carried far by the model's dynamics; a signed accounting of per-step truncation residuals propagated through the linearized dynamics reconstructs the endpoint error's direction and regional structure. If correct, this separates 'where error is injected' from 'where it lands' and shows both are readable from the model, with the injection landscape modifiable by a training penalty.","feed_headline":"Under 10% of a perturbation's effect stays in its source region","feed_subtitle":"Per-region velocity variation predicts where few-step solvers inject error; propagating residuals shows where it lands.","key_machinery":"The argument rests on two objects. Flow Complexity (FC) is a per-region total variation: for flow models it is the sum of the magnitudes of the learned velocity field's changes along the trajectory, and for diffusion models the corresponding variation of the prediction field. FC is the quantity the Euler truncation bound ties to local injection: each region's accumulated variation bounds the magnitude of the local truncation residual it injects. The second object is the state-transition operator Phi of the linearized dynamics, which carries each injected residual to the endpoint; Proposition 1 bounds final per-region error by a Phi-weighted sum of all regions' FC values. The reconstruction p","core_discovery":"The paper's central claim is that few-step integration error in generative ODEs has a learned, spatially structured source-and-transport story. A per-region statistic, Flow Complexity, defined as the accumulated variation of the model's velocity (or prediction) field along a fine reference trajectory, bounds the truncation the solver injects in each region, and a leading-order bound (Proposition 1) relates final per-region error to Flow Complexity values carried through the state-transition operator of the linearized dynamics. Empirically, Flow Complexity predicts per-region error within images at Spearman rho 0.57-0.70 on five 256px models, but only partially: a perturbation protocol shows","pith_inferences":["Editorial: If injected error is transported nonlocally, the natural next test is an adaptive step allocator that assigns budget to projected 'landing' regions (computed from a cheap propagated-residual estimate) rather than to local FC; the paper does not offer such a sampler, but its transport kernel suggests it.","Editorial: The direction-, sign-, and timing-dependence of the reconstruction suggests that a penalty on signed cumulative variation or a horizon-weighted variant might straighten trajectories more efficiently than the magnitude penalty the paper tests; the paper reports such variants predict error as well as or better than FC, so a training objective built on them is a plausible extension.","Editorial: The prediction horizon is model-dependent (early for flow models, late for diffusion), which suggests that 'when difficulty becomes readable' is itself learned; an online probe that detects the error-determining window during sampling could turn the diagnostic into a stopping rule, though the paper does not establish that the cheap-solve signal is strong enough to drive such a rule.","Editorial: The paper's finding that refining the lowest-FC regions improves FID while refining the highest-FC regions reduces integration error implies numerical integrability and perceptual quality are different objectives; a testable extension is to measure whether propagation-aware correction changes FID differently than injection-aware correction."],"forward_implications":["If the paper is right, per-region difficulty maps are not noise: they are stable, model-specific structures that emerge during training, so a solver or sampler could treat them as part of the model's signature.","Early-window Flow Complexity, even computed only from the first three steps of a cheap solve, localizes a meaningful share of final error (about a quarter, against a 14.8% random baseline in the paper's protocol-frozen tests), so adaptive refinement can in principle start before the trajectory is complete.","Because a region's final error is dominated by error transported in from other regions, improving few-step sampling by refining only the locally hardest regions will be incomplete; an allocation rule must account for propagation through the linearized dynamics.","A training-time penalty on Flow Complexity reduces both injected error and coarse-solve error and improves few-step FID at matched optimization steps (57-65% integration-error reduction; about 7% FID gain at NFE 6 on ImageNet), so the difficulty landscape is something training can change.","The paper's own compute-matched analysis shows the training penalty wins at moderate scale but not at ImageNet scale when the extra forward per step is charged, so the mechanism's production value remains open."],"fun_headline_variants":["Solver error spreads: <10% stays at source","Velocity variation predicts where solver error lands","Few-step error is a transport problem, not local","Under 10% of solver error stays put; rest travels"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The reconstruction's per-step truncation residuals are evaluated along the near-exact reference trajectory, not along the path the cheap solver actually visits; if the cheap trajectory diverges from the reference, the residuals injected during real sampling differ from the ones being propagated.","fun_headline_variants_meta":{"raw":{"variants":["Solver error spreads: <10% stays at source","Velocity variation predicts where solver error lands","Few-step error is a transport problem, not local","Under 10% of solver error stays put; rest travels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1178,"prompt_tokens":830,"completion_tokens":348,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":285}},"tokens_in":574,"tokens_out":348,"duration_ms":4396,"temperature":1.0,"reasoning_tokens":285,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T21:56:44.670628+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the divergence between an NFE-6 Euler path and the fine reference path on the paper's own models, and recompute the reconstruction with residuals evaluated along the cheap path instead of the reference path: if the endpoint cosine (0.81-0.87) and per-region rho (0.80-0.82) drop to the no-propagation level, the transport account describes reference-path error rather than the error experienced by the actual solve.","supporting_citations":[],"review_version":1}