{"id":"90685727-d2e8-4147-a890-2677d43d75a0","arxiv_id":"2507.09888","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A flow-matching model with a neural-operator velocity field is proposed to forecast time series by transporting distributions over continuous functions, reporting top average rank on eight benchmarks.","lead":"NeuTSFlow treats time series forecasting as moving probability mass between families of continuous functions rather than mapping discrete points. It couples flow matching with neural operators and reports the best average rank on eight standard benchmarks, though one key inference step in the paper is internally inconsistent.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 3.1 gives two incompatible definitions of the conditional velocity (f−h vs. f); the stated ODE inference then yields h+f, so the forecasting mechanism depends on an unstated conversion.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the two definitions of the conditional velocity in Section 3.1 are mutually inconsistent, and the stated ODE inference does not produce the target function. My reading of the full text confirms this. I also note two reinforcing ambiguities: Eq. (3) describes prediction as a single evaluation rather than an ODE solve, and the ablation 'w/o flow matching' cannot be interpreted until the actual inference recipe is known. The high-level idea of transporting a distribution over functions is coherent and the experimental setup is standard, but the central mechanism as written is not self-consistent. Because the appendix with the detailed inference procedure is absent from the reviewed text, this is not merely a typographical issue; the paper does not yet specify how a reader would reproduce the claimed endpoint. The concern is internal to the manuscript, not a disagreement with an external consensus, and it directly affects the validity of the reported flow-matching contribution. I therefore see no reason to change the reader's REJECT verdict; a local fix (e.g., train on f − h, or explicitly use u_theta − h as the ODE velocity) would likely resolve it.","tokens_in":12895,"tokens_out":4608,"duration_ms":55717,"concrete_test":"Implement the inference procedure exactly as described in Section 3.1-3.2 on a synthetic pair (h, f): instantiate or train the model so that u_theta(t, g) outputs f (the reparameterized target), set g_0 = h, and solve dg_t/dt = u_theta(t, g_t) with a standard ODE solver to t = 1. If the endpoint equals f, the solver must be using an additional conversion not present in the equations; if the endpoint equals h + f, the inconsistency is confirmed. Equivalently, inspect the Appendix A.3 pseudocode or released code: the velocity passed to the ODE solver must be u_theta − h (or the model must be trained on f − h) for the endpoint claim to hold. If no ODE solver appears in the code and the model is evaluated one-shot as in Eq. (3), that itself refutes the flow-matching inference narrative.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (2) defines the conditional velocity as v^{f,h}_t(g) = f − h for the straight-line path g_t = t f + (1 − t) h. The paragraph on Target Reparameterization then redefines v^{f,h}_t(g) = f, and the loss in Eq. (6) is regressed against this redefined velocity. These two definitions cannot both describe the same path. If the trained model u_theta is used as the reparameterized velocity f, integrating dg_t/dt = u_theta(t, g_t) from g_0 = h gives g_1 = h + f, not f. The paper's statement that solving the ODE 'until t = 1, at which point f is obtained' is therefore not a consequence of the equations as written. To be internally consistent, the authors must either keep Eq. (2) as the true ODE velocity and convert the model output f into f − h during inference (an unstated step), or drop the ODE and present the method as direct functional regression. Moreover, the 'complete prediction process' in Eq. (3) is written as a single function evaluation, Y = N^{-1}[u_theta(N(H), t, G_t)], with no ODE solver shown, while the prose describes ODE integration; the exact inference recipe is ambiguous. The ablation labeled 'w/o flow matching' cannot be interpreted without resolving this ambiguity: if both the full model and the ablation are evaluated by direct evaluation of u_theta, the reported contribution of flow matching is not actually measured. Appendix A.3, said to contain the training and inference details, is not present in the reviewed text, so the required conversion cannot be verified from the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NeuTSFlow, a framework that reframes time series forecasting as learning a transition between probability measures over continuous functions, rather than a mapping between discrete sequences. The method constructs a Gaussian measure path between historical and target functions, defines a conditional velocity field, and learns a neural-operator velocity-field model via a flow-matching loss. The authors claim that the resulting model is a continuous generalization of DLinear and report state-of-the-art results on eight benchmark datasets across three tasks: conventional forecasting, time series super-resolution, and cross-resolution temporal learning. The paper also provides propositions about marginalization, loss equivalence, and function-to-function mapping, and it includes ablation studies purporting to show the importance of the neural operator and flow-matching components.","tokens_in":13143,"tokens_out":5159,"duration_ms":61289,"significance":"If the framework were internally consistent, the function-family perspective and the use of neural operators as velocity fields would be a potentially interesting contribution to time series forecasting. The empirical evaluation is broad, covering three tasks and eight datasets, and the paper makes falsifiable claims about the value of the neural-operator and flow-matching components. However, the central flow-matching construction is inconsistent as written: the conditional velocity is defined in two incompatible ways, and the stated inference procedure does not follow from the equations. Because this inconsistency concerns the core mechanism of the method, the significance of the reported results cannot be assessed without substantial revision. The paper also lacks the appendix containing the training and inference details, which further impedes verification.","major_comments":[{"comment":"Eq. (2) defines the conditional velocity as v^{f,h}_t(g) = f - h for the straight-line path g_t = t f + (1 - t) h, but the following paragraph redefines it as v^{f,h}_t(g) = f. These definitions are incompatible. If the learned model u_theta approximates v^{f,h}_t(g) = f, then solving the ODE dg_t/dt = u_theta(t, g_t) from g_0 = h yields g_1 = h + f, not f. The claim in Section 3.2 that 'this process continues until t = 1, at which point f is obtained' is therefore not a consequence of the equations as written. The paper does not state any conversion from the model output f to the ODE velocity f - h during inference. This is a load-bearing issue because the entire forecasting mechanism depends on the velocity field used for integration.","section":"Section 3.1, Eq. (2) and 'Target Reparameterization for Functional Forecasting'"},{"comment":"The 'complete prediction process' is written as a single function evaluation, Y = N^{-1}[u_theta(N(H), t, G_t)], with no ODE solver appearing in the equation, while the prose describes solving an ODE from t = 0 to t = 1. The exact inference recipe is therefore ambiguous. This ambiguity also undermines the ablation labeled 'w/o flow matching' in Table 4: if both the full model and the ablation are evaluated by direct evaluation of u_theta, then the reported contribution of flow matching is not actually measured. The referenced Appendix A.3, which is said to contain the training and inference details, is not present in the reviewed text, so the required procedure cannot be verified.","section":"Section 3.2, Eq. (3)"},{"comment":"With the reparameterized velocity v^{f,h}_t(g) = f, the conditional loss in Eq. (6) is exactly a direct regression of the model output u_theta(t, g_t) to the target function f at every time t. Because h is fixed and f is independent of g_t, the optimal model can ignore g_t entirely, and the ODE integration contributes nothing beyond the direct prediction of f. The paper provides no experiment that separates the effect of ODE integration from direct functional regression, so the claim in Table 4 that 'flow matching' is the second-most critical component is not supported by the evidence presented. A consistent formulation would require either regressing on f - h and using that as the ODE velocity, or explicitly deriving how the model output f is converted into the velocity field during inference, and then demonstrating that the integration step adds predictive value beyond a single evaluation.","section":"Section 3.1, Eq. (6), and Table 4"}],"minor_comments":[{"comment":"The manuscript contains numerous typos and unfinished edits, for example 'to guaranteeensure', 'align witmatch', and 'the predictedion data' in Section 3.2. These should be corrected.","section":"Throughout"},{"comment":"Eq. (1) writes g_t = mu^{f,h}_t = N(tf + (1 - t)h, C_t), but g_t is subsequently used as a function rather than a measure. This notational conflation is confusing and should be clarified.","section":"Eq. (1)"},{"comment":"Reference [3] is cited as Autoformer but the title given, 'Searching transformers for visual recognition', corresponds to a different paper. The citation and bibliography entry need to be corrected.","section":"References"},{"comment":"The experimental tables report only MSE values without standard deviations or statistical significance tests. Given that NeuTSFlow is claimed to rank first on almost every dataset and task, the absence of variance estimates makes it difficult to assess whether the improvements are meaningful.","section":"Tables 1-3"}],"recommendation":"reject","confidential_remarks":"The core flow-matching construction in Section 3.1 is internally inconsistent, and the missing Appendix A.3 makes it impossible to verify the training and inference details. If the authors were to reformulate the method as a direct functional regression model with neural operators, the empirical results might still be of interest, but that would be a substantially different paper. As written, the central claim that flow matching learns a path between function families is not established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the function-family framing is worth taking seriously, but the paper has a load-bearing inconsistency in Section 3.1 that breaks the stated inference. Eq. (2) defines the conditional velocity as f−h; two paragraphs later, the target reparameterization redefines it as f. If the trained velocity field is f, integrating the ODE from h gives h+f at t=1, not f. The claim that solving until t=1 yields f is not a consequence of the equations as written. The appendix might clarify, but the main text contradicts itself and the burden is on the authors.\n\nWhat is genuinely new: the idea of treating discrete series as samples from a measure on continuous functions and transporting historical to future function families with a neural-operator velocity field is a fresh and sensible angle. The super-resolution and cross-resolution tasks are a natural testbed, and the reported wins across all datasets are impressive if they hold. The architecture—normalization, spectral decomposition, dimension expansion, spectral temporal kernel—is reasonably motivated, and the ablation at least suggests the components matter.\n\nSoft spots beyond the central inconsistency: the inference recipe is ambiguous between Eq. (3), a direct function evaluation, and the prose describing ODE integration. That makes the 'w/o flow matching' ablation hard to interpret: if both the full model and ablation are evaluated as direct mappings, the contribution of flow matching is not actually measured. The propositions are mostly standard flow-matching facts or empty statements; Proposition 3.2 asserts an architectural property with no derivation in the main text. No code, seeds, or error bars are provided, so the empirical claims are hard to verify.\n\nI agree with the reader's negative verdict and the stress-test note; the issue is real and central. That said, the fix looks local (use f−h as velocity, or present the method as direct functional regression), and the framing deserves a serious referee. I'd send it to review, not desk-reject, with the expectation of major revision and a request for code and a consistent derivation.","headline":"A promising function-family framing with a load-bearing inconsistency: the conditional velocity is defined as f−h then redefined as f, so the ODE inference yields h+f and the paper's central claim does not follow as written.","tokens_in":13807,"tokens_out":4350,"would_cite":false,"duration_ms":43405,"reading_group":"yes","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 time series forecasting is best modeled as a continuous flow between probability measures over functions, and reports that a neural-operator velocity field achieves the best average rank across eight benchmarks.","keywords":["time series forecasting","flow matching","neural operator","function space","probability measure","continuous functions","super-resolution","cross-resolution learning"],"falsifier":"Check the implementation to see whether inference actually time-marches an ODE through the neural operator or produces the forecast in a single forward pass; if no ODE is solved, the flow-matching stage is not the mechanism behind the reported accuracy.","tokens_in":12589,"feed_emoji":"📈","tokens_out":11968,"duration_ms":109327,"temperature":0.7,"pith_summary":"This paper argues that a time series is best understood not as a discrete sequence but as a noisy sample from a family of continuous functions, and that forecasting is the task of transporting the probability measure over historical functions to the probability measure over future functions. It introduces NeuTSFlow, a flow-matching model whose velocity field is a neural operator, so that solving an ordinary differential equation in function space carries a historical function into a future function. On eight standard datasets, the authors report the best average rank for conventional forecasting and first place on every dataset for time-series super-resolution and cross-resolution learning. The significance of the claim is that modeling function-level structure, rather than point-by-point dependencies, could be the more faithful and transferable basis for forecasting.","feed_headline":"Function-space flow beats eight forecasting benchmarks","feed_subtitle":"NeuTSFlow models prediction as a flow between function distributions and beats point-based rivals on average.","key_machinery":"The central object is the marginal velocity field $u_\\theta^t: \\mathcal{F} \\times [0,1] \\to \\mathcal{F}$, a time-dependent neural operator that defines the flow ODE $\\partial_t \\phi_t(g) = u_\\theta^t(t, \\phi_t(g))$ from a historical function $h$ to a future function $f$. A neural operator is a model that maps between infinite-dimensional function spaces rather than between fixed-length vectors. The velocity field is trained by the conditional flow-matching loss $L_{\\text{con}}(\\theta) = \\mathbb{E}_{t, f \\sim \\mu_F, h \\sim \\mu_H} \\| v^{f,h}_t(g) - u_\\theta^t(g) \\|^2$, using the straight-line path $g_t = t f + (1-t)h$, and is implemented with a spectral kernel in the Fourier domain that performs the role of the integral operator $\\kappa(y,x)$ in neural-operator theory. This construction is what lets the model transport whole distributions of functions instead of predicting isolated future points.","core_discovery":"The central claim is that the forecasting map $M: \\mu_H \\to \\mu_F$ can be learned by constructing a conditional path of measures $\\mu_{f,h}^t = \\mathcal{N}(t f + (1-t)h, C_t)$ between paired historical and future functions $h$ and $f$, and by training a neural operator to reproduce the conditional velocity of that path. The paper shows that as $C_t \\to 0$ the conditional velocity is $f - h$, the straight-line transport direction, and then reparameterizes the target so the model regresses directly to $f$. The learned marginal velocity field $u_\\theta^t$ is built from normalization, spectral decomposition, dimension expansion, and frequency-domain temporal learning, and Proposition 3.2 states that it is a neural operator and a continuous-time version of DLinear. The paper's empirical claim is that this function-family approach ranks first on all eight datasets for super-resolution and cross-resolution tasks and achieves the best average rank on conventional forecasting.","pith_inferences":["The paper leaves implicit how the two conditional velocities in Section 3.1, $f-h$ and $f$, are reconciled during ODE integration; a clarifying extension would be to state that the predicted $f$ is converted to $f-h$ before solving the flow.","A testable extension is to run inference both with full ODE integration and with a single direct prediction; if the outputs are identical, the flow-matching stage is not the source of the reported gains.","The function-family view suggests a natural probabilistic-forecasting extension: sample several functions from the transported measure $\\mu_F$ to obtain calibrated prediction intervals, which the paper does not develop."],"forward_implications":["If discrete time series are noisy samples of continuous functions, then a model trained at one sampling resolution can predict at another, which is what the paper's super-resolution and cross-resolution experiments show.","The claim that $u_\\theta^t$ is a continuous-time DLinear connects flow matching to classical linear forecasting, implying that linear models can be read as discretizations of a function-space flow.","Because the flow transports a full probability measure over functions, the same framework can in principle output a distribution of plausible futures rather than a single point forecast.","Solving the flow ODE from an arbitrary historical function gives a path to extrapolate beyond trained forecast horizons without retraining the model."],"supporting_citations":[{"why":"Supplies the functional flow-matching framework and the marginalization construction the paper builds its measure path on.","marker":"[9]"},{"why":"Provides the conditional flow-matching loss and the gradient-equivalence result used to train the velocity field.","marker":"[14]"},{"why":"Gives the rectified-flow straight-line velocity used in the paper's measure path.","marker":"[16]"},{"why":"Supplies the Fourier neural operator parameterization behind the spectral temporal learning module.","marker":"[13]"},{"why":"Establishes neural operators as learnable maps between function spaces, the basis for treating the velocity field as a neural operator.","marker":"[12]"},{"why":"DLinear is the classical baseline that Proposition 3.2 claims the model is the continuous-time version of.","marker":"[30]"},{"why":"Justifies the use of Gaussian measure paths in infinite-dimensional spaces through absolute continuity of Gaussian measures.","marker":"[21]"},{"why":"Supports the function-space flow perspective on learning conditional paths between measures.","marker":"[23]"}],"fun_headline_variants":["Function-family flow outranks point models on 8 tasks","Forecast as flow between continuous function families","NeuTSFlow: neural operator flow for time series","From discrete points to continuous functions for prediction","Flow matching on function spaces wins forecasting benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method depends on the conditional velocity field being defined consistently during training and inference; Section 3.1 first defines it as $f-h$ and then redefines it as $f$, so the claim that solving the ODE from $h$ yields $f$ needs an unstated conversion step to hold.","fun_headline_variants_meta":{"raw":{"variants":["Function-family flow outranks point models on 8 tasks","Forecast as flow between continuous function families","NeuTSFlow: neural operator flow for time series","From discrete points to continuous functions for prediction","Flow matching on function spaces wins forecasting benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000209,"raw_usage":{"total_tokens":1416,"prompt_tokens":964,"completion_tokens":452,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":380}},"tokens_in":580,"tokens_out":452,"duration_ms":5815,"temperature":1.0,"reasoning_tokens":380,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:45:55.645938+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check the implementation to see whether inference actually time-marches an ODE through the neural operator or produces the forecast in a single forward pass; if no ODE is solved, the flow-matching stage is not the mechanism behind the reported accuracy.","supporting_citations":[{"cited_title":"Absolute continuity of Brownian bridges under certain gauge transformations","cited_arxiv_id":"1103.4822","evidence_quote":"Justifies the use of Gaussian measure paths in infinite-dimensional spaces through absolute continuity of Gaussian measures."}],"review_version":1}