{"id":"d7c9b857-fea9-4565-bdae-e7a8b13e7f46","arxiv_id":"2508.18565","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SPF trains a one-step forecaster on a mixture of ground-truth and self-generated states, improving long-term rollout accuracy on two fluid benchmarks with constant memory use.","lead":"This paper proposes Stochastic PushForward (SPF), a training method that mixes a model's own predictions with real data during one-step training to improve long-horizon forecasting. Experiments on fluid dynamics benchmarks suggest lower long-term error and memory use than autoregressive training, but the comparison is incomplete on one benchmark.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 appends (rather than replaces) predicted states to Dδ on every NUI update, so the supplementary dataset grows without bound and the claimed stable memory / constant-cost behavior is not what is implemented.","rationale":"The paper's stated contribution is two-fold: better long-term accuracy than ATF/PF and lower, δ-independent memory. The reader's weakest assumption targets the accuracy side (Dδ states must match rollout states). That concern is real but generic and can be addressed by a distribution analysis. I found a more immediate, checkable issue on the efficiency side: the update rule in Algorithm 1 as written appends newly generated states to Dδ rather than replacing the old snapshot. Every NUI cycle therefore grows the dataset, which contradicts the 'memory usage stable' claim and causes training-time growth; it also mixes stale predictions into the training distribution. This is not a matter of interpreting experimental plots; it can be settled by reading the released code and logging dataset growth over epochs. Because the concern is falsifiable and addressable rather than fatal, I keep the reader's CONDITIONAL verdict.","tokens_in":16364,"tokens_out":10553,"duration_ms":110852,"concrete_test":"Inspect the released GitHub code to see whether Dδ is cleared and regenerated at each update (Dδ = new predictions) or accumulated via an append operation. Then run a controlled shallow-water sweep with NUI=1 for 20 sequential epochs, logging dataset length, per-epoch wall time, and peak CPU/GPU memory. If dataset length and per-epoch time grow with the number of updates while peak memory remains constant, the stable-memory claim and Table 3 do not describe the algorithm actually trained, and the SPF accuracy comparison should be re-run with a bounded or replacement-based Dδ.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 and Algorithm 1 define the supplementary-dataset update as 'Dδ ← Dδ ∪ {~ηt+δ}' inside the 'if (epoch mod NUI) == 0' branch, and the comment says 'Append predicted sequence to Dδ.' This is an append, not a reset: every update adds (T−δ) newly generated δ-step-ahead states, and because the model has changed since the previous update, these are not duplicates that a set union would collapse. Thus Dδ grows without any specified bound or replacement rule. The paper's central efficiency claim—constant memory independent of δ, specifically 689 MB for δ=2 and δ=3 in Table 3—assumes a fixed-size training set. An accumulated Dδ increases CPU memory, raises per-epoch data-transfer and training cost, and preserves stale predictions from earlier checkpoints in the training mix. That stale data also threatens the accuracy mechanism: the model is no longer trained on a 'periodically updated' snapshot of current-model predictions, so the transfer from Dδ to actual rollout states cannot be assessed. Either the pseudocode is inaccurate or the memory/efficiency comparison is overstated; the current text does not let the reader reproduce the SPF update.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Stochastic PushForward (SPF), a training framework for long-term autoregressive forecasting that combines one-step-ahead training with a stochastically sampled mixture of ground-truth and model-predicted inputs. SPF builds a supplementary dataset by unrolling a frozen copy of the surrogate model for δ steps from ground-truth states, refreshes it periodically, and trains a one-step surrogate using a Bernoulli acquisition rule with a weighting coefficient α. Multi-step predictions are precomputed between epochs, so gradients are never backpropagated through an unrolled sequence. The framework is evaluated on 2D Burgers' equation and a shallow-water benchmark using a CAE-LSTM architecture, with comparisons against a basic LSTM, autoregressive training (ATF), and PushForward (PF), plus additional experiments on data scarcity, time-step extrapolation, noisy inputs, energy consistency, a physics-constrained variant, hyperparameter sensitivity, and GPU memory usage. The paper claims higher long-term accuracy and lower memory than autoregressive approaches.","tokens_in":16645,"tokens_out":6383,"duration_ms":57725,"significance":"If validated, SPF fills a practical niche: it obtains multi-step robustness without backpropagating through multiple unrolled steps, while preserving short-term accuracy because the training loss remains one-step. The paper's strengths are that it provides a concrete algorithm, a public code repository, a range of diagnostic experiments (limited data, extrapolation, noisy inputs, physics-constrained variant), and a direct comparison with ATF and PF on the shallow-water benchmark. The central mechanism is plausible and the shallow-water results are encouraging. However, the current manuscript does not yet fully support the headline claims: the Burgers benchmark lacks the ATF/PF comparison, the SSIM truncation hides part of the rollout behavior, and the unbounded append rule in Algorithm 1 conflicts with the stable-memory claim. These issues are fixable within the paper's scope.","major_comments":[{"comment":"In Algorithm 1, the supplementary-dataset update is implemented as 'Dδ ← Dδ ∪ {~ηt+δ}' inside the 'if (epoch mod NUI) == 0' branch, with the comment 'Append predicted sequence to Dδ.' This is an appending operation, not a replacement or reset. Because the model changes between updates, successive updates add new predicted states that are not duplicates, so |Dδ| grows without bound as training proceeds. This directly contradicts the claim in §5.7 that SPF maintains constant RAM usage of 689 MB independent of δ, since an unbounded Dδ increases CPU-side storage and, depending on the data-transfer strategy, per-epoch data-transfer and training cost. It also preserves stale predictions from earlier checkpoints in the training mix, weakening the paper's mechanism that Dδ represents the current model's rollout states. The pseudocode must either be corrected to a replace/refresh operation (matching the memory claims) or the memory and training-cost analysis must account for the accumulated supplementary dataset. This is load-bearing because both the headline efficiency claim and the accuracy mechanism depend on the actual update rule.","section":"Algorithm 1, Section 3.4"},{"comment":"The Burgers' equation experiments compare only the basic one-step LSTM with SPF2/SPF3; neither ATF nor PF is evaluated on this test case. The paper's claims in the abstract and introduction that SPF outperforms 'autoregressive approaches' are therefore only directly supported by the shallow-water experiments. The Burgers results should either include ATF/PF baselines with the same depth settings, or the claims should be restricted to the shallow-water benchmark.","section":"Section 4 (Burgers' equation)"},{"comment":"The text states that 'if SSIM falls below 0.8, the subsequent data points are not plotted, as the results below the threshold are deemed non-informative.' This truncation can hide large accumulated errors: for models whose SSIM drops below threshold early, such as the basic LSTM, the plotted accumulated-error curve stops, and the visual comparison ends at the truncation point rather than at the full horizon. The reported metric comparisons are therefore incomplete. Please report full-length curves or, at minimum, state the truncation explicitly in every affected figure and provide the step at which each model crosses the threshold.","section":"Section 5.2.1, Fig. 9"},{"comment":"No random seeds, number of independent runs, or error bars are reported for any experiment. The claim in §5.2.1 that 'SPF models demonstrate less error and uncertainty' appears to refer to standard deviations, but the corresponding curves are not shown with error bands, and the number of seeds is not given. In addition, key training hyperparameters (LSTM architecture, optimizer settings, learning rate, batch size, Ninit, Nepoch, NUI, and the chosen p and α for each reported result) are not stated in the main text. The repository link is helpful, but the paper should include the numerical settings needed to reproduce Tables 1–3 and Figures 6–16 without inferring them from code.","section":"Section 5 (experimental setup)"},{"comment":"The memory comparison is ambiguous. Table 3 is headed 'RAM Usage,' but §5.7 states that 'Both ATF and SPF initially store all data on the CPU and transfer it to the GPU only for training.' It is not clear whether the reported numbers are GPU memory, CPU RAM, or total memory, nor whether the supplementary dataset Dδ storage is included in the SPF figures. Since the central efficiency claim is that SPF keeps memory 'stable regardless of δ,' the measurement procedure must be stated precisely. If the reported 689 MB excludes the CPU-side Dδ storage, Table 3 understates the true memory footprint when Dδ is accumulated as in Algorithm 1.","section":"Table 3 and Section 5.7"}],"minor_comments":[{"comment":"The word 'recrusive' should be 'recursive.'","section":"Section 2"},{"comment":"The phrase 'singe-step-ahead prediction' should be 'single-step-ahead prediction.'","section":"Section 3.1"},{"comment":"Since I_t is Bernoulli, the expression 'ηI_t = I_t·η^{D1}_t + (1−I_t)·η^{Dδ}_t' is a selection between two states rather than a convex combination; presenting it as a conditional selection would be clearer.","section":"Eq. (13)"},{"comment":"The statement that 'the performance of PF models deteriorates as δ increases' is not connected to the quantitative results in Fig. 9, where PF3 does not uniformly underperform PF2; please clarify or qualify this claim.","section":"Section 5.2.2, Fig. 10"},{"comment":"The phrase 'To access whether the model meets this criterion' should be 'To assess whether the model meets this criterion.'","section":"Section 5.4"},{"comment":"The notation 'length(D)' is not explicitly defined; since the dimensionality of D depends on whether Dδ is reset or appended, the algorithm should define it precisely.","section":"Algorithm 1"},{"comment":"It should be stated whether the reported MSE and SSIM values are from a single run or averaged over multiple seeds.","section":"Tables 1–2"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the central idea is worth pursuing, but the unresolved gap between Algorithm 1's append semantics and the constant-memory claim needs to be fixed before publication. I also recommend asking the authors to verify that the GitHub repository matches the revised pseudocode and to provide the actual hyperparameter values used for Tables 1–3."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"SPF is a plausible low-memory alternative to backprop-through-time for neural surrogate models, but the paper's own Algorithm 1 contradicts its central constant-memory claim.\n\nThe core idea is clean: keep one-step-ahead training, but every few epochs unroll the frozen model to build a supplementary dataset of predicted states, then mix those with ground truth via Bernoulli sampling. On the shallow-water benchmark it beats ATF and PF on long-term MSE/SSIM, and Table 3 shows constant 689 MB for δ=2 and 3 versus growing usage for ATF/PF. The work is clearly written, the code is public, and the conclusion honestly lists the added complexity of the method.\n\nThe main novelty is the precompute-between-epochs trick and the stochastic acquisition rule. It is close to Vlachas & Koumoutsakos [34], but the packaging genuinely differs: a periodically updated, stochastically sampled supplementary dataset rather than fusing predictions during the backward pass. The limited-data experiments are a useful addition.\n\nThe soft spots are real. The biggest is the algorithm itself. In Algorithm 1, the update is 'Dδ ← Dδ ∪ {η̃_{t+δ}}' with the comment 'Append predicted sequence to Dδ.' That is an append, not a reset or a bounded replacement. Dδ grows without bound every NUI epochs. The paper's memory claim—'SPF consistently maintains RAM usage at 689MB for both δ=2 and δ=3'—only makes sense if Dδ is a fixed-size dataset. Either the pseudocode is inaccurate (and the implementation actually caps or replaces Dδ) or the memory comparison is overstated. The current text does not let a reader reproduce SPF, and this is load-bearing for the paper's headline efficiency claim.\n\nTwo other gaps: Burgers only compares SPF to a basic LSTM, not ATF/PF, so the 'superior to autoregressive' conclusion rests on shallow water alone. And there are no error bars or seeds reported; the baseline hyperparameters (rollout schedules, loss weights) are under-specified. The SSIM truncation at 0.8 is acceptable, but they should say how many runs fall below threshold.\n\nSeparately, the paper cites [34] but never explains what SPF changes relative to it. A short comparison paragraph would settle that.\n\nThis deserves a serious referee—the idea is plausible and the shallow-water results are encouraging—but the manuscript needs major revision before it is acceptable. The dataset update rule must be clarified, Burgers needs the missing baselines, and variability needs to be reported. As written, I would not cite the constant-memory claim.","headline":"SPF is a plausible low-memory alternative to backprop-through-time for neural surrogates, but the paper's own Algorithm 1 contradicts its central constant-memory claim.","tokens_in":17116,"tokens_out":3048,"would_cite":false,"duration_ms":27629,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper's central claim is that training on a stochastic mixture of ground-truth and model-generated states can replace multi-step backpropagation for long-term forecasting, improving accuracy at constant GPU memory.","keywords":["Stochastic PushForward","long-term prediction","autoregressive training","supplementary dataset","Burgers' equation","shallow water equations","memory-efficient training","LSTM surrogate"],"falsifier":"Train SPF, then record both the states in the supplementary dataset and the states the final model visits during long held-out rollouts, and compute a distributional distance between the two sets in latent space. If SPF still improves long-horizon accuracy when that distance is large, the stated correction mechanism is not the one doing the work; if long-horizon error tracks that distance, the premise is load-bearing.","tokens_in":16170,"feed_emoji":"🌊","tokens_out":8747,"duration_ms":79475,"temperature":0.7,"pith_summary":"This paper proposes Stochastic PushForward (SPF), a training framework that keeps one-step-ahead training while still teaching the model about multi-step errors. SPF periodically unrolls the current model from ground-truth states, stores those predictions in a supplementary dataset, and stochastically mixes them with real data at every update. The claim is that this balances short- and long-term accuracy without the GPU memory growth of autoregressive training (ATF) or PushForward (PF). On the Burgers' equation and shallow water benchmarks, SPF reports lower accumulated error and higher structural similarity over long horizons, while keeping memory at a constant 689 MB for depths δ=2 and δ=3.","feed_headline":"Training on mixed real and predicted data improves long-term forecasts","feed_subtitle":"SPF injects model-generated states into one-step training, beating autoregressive and PushForward baselines on fluid benchmarks.","key_machinery":"The load-bearing object is the supplementary dataset Dδ of model-generated latent states, refreshed every NUI epochs, together with the stochastic acquisition rule It~Bernoulli(p) that decides whether the input for the one-step loss comes from ground truth or from Dδ. The target is always the ground-truth next state, and a coefficient α down-weights the supplementary samples. Because the δ-step composition fδ is evaluated only to generate data between epochs, the training loop stays one-step-ahead and memory use does not grow with δ.","core_discovery":"SPF's central claim is that multi-step learning can be achieved without multi-step backpropagation by training on a mixture of ground-truth and self-generated states. The paper defines a supplementary dataset Dδ built by applying the composed predictor fδ to ground-truth inputs, then samples each training input from D1 or Dδ according to a Bernoulli draw with probability p, always targeting the true next state, and weights supplementary samples by α. Multi-step predictions are precomputed between epochs, so no unrolled sequence occupies GPU memory during gradient computation. Experiments on Burgers' equation and the shallow water system report that SPF outperforms the base LSTM, ATF, and PF on accumulated error and SSIM over long horizons, preserves short-term accuracy, remains robust under noisy and sparse data, and keeps RAM at 689 MB independent of depth δ.","pith_inferences":["An untested extension is to anneal the sampling probability p from mostly ground-truth early in training to mostly predicted states later, turning SPF into a curriculum that may reduce sensitivity to the fixed hyperparameter values the paper reports.","Because memory does not scale with δ, SPF could in principle train very deep rollouts for transformer-based latent surrogates where ATF and PF storage becomes prohibitive; a chaotic test problem would show whether the claimed balance extends beyond the two smooth benchmarks.","The supplementary dataset is a form of experience replay, so prioritizing predicted states with the largest deviation from ground truth could target exactly the rollouts that diverge, a variant not explored in the paper."],"forward_implications":["On both the Burgers' equation and shallow water benchmarks, SPF lowers accumulated error and raises SSIM over long horizons relative to the base LSTM, ATF, and PF.","GPU memory stays at 689 MB for depths δ=2 and δ=3 in the shallow water test, while ATF and PF memory grows with depth, making SPF's cost independent of rollout depth.","SPF preserves short-term one-step accuracy because ground-truth samples remain in the stochastic mix, and it maintains higher SSIM than ATF/PF under noisy inputs.","SPF remains the strongest method when training data is reduced to 50%, 30%, 10%, and 5% of the full set, and it can be combined with an energy-regularization term to reduce accumulated error further."],"supporting_citations":[{"why":"Introduces the PushForward training method that SPF extends and serves as the direct baseline for accuracy and memory comparisons.","marker":"[36]"},{"why":"Describes the autoregressive training framework (ATF) that SPF identifies as memory-intensive and uses as a comparison baseline.","marker":"[33]"},{"why":"Formulates learning from predictions for long-term spatiotemporal forecasting, setting the conceptual stage for training on the model's own outputs.","marker":"[34]"},{"why":"Supplies the shallow water benchmark used for the main ATF/PF/SPF comparisons.","marker":"[37]"},{"why":"Demonstrates the CAE-LSTM reduced-order architecture that SPF uses to compress solution fields before training.","marker":"[23]"},{"why":"Defines the LSTM surrogate model used as the dynamics predictor in all training frameworks.","marker":"[45]"}],"fun_headline_variants":["SPF mixes real and predicted data to boost long-term forecasts","Mix real and predicted states to get better long-term forecasts","SPF cuts memory and improves long-term predictions","Better long-term forecasts with mixed training data","Training on real and model-generated data extends forecast accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the model's own earlier predictions, generated by unrolling from ground-truth states and stored in the supplementary dataset, resemble the states the model will actually meet during long test-time rollouts, so that a one-step loss on those predicted states teaches it to reduce compounding errors.","fun_headline_variants_meta":{"raw":{"variants":["SPF mixes real and predicted data to boost long-term forecasts","Mix real and predicted states to get better long-term forecasts","SPF cuts memory and improves long-term predictions","Better long-term forecasts with mixed training data","Training on real and model-generated data extends forecast accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000804,"raw_usage":{"total_tokens":3494,"prompt_tokens":868,"completion_tokens":2626,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":2550}},"tokens_in":484,"tokens_out":2626,"duration_ms":17610,"temperature":1.0,"reasoning_tokens":2550,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:56:03.047052+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SPF, then record both the states in the supplementary dataset and the states the final model visits during long held-out rollouts, and compute a distributional distance between the two sets in latent space. If SPF still improves long-horizon accuracy when that distance is large, the stated correction mechanism is not the one doing the work; if long-horizon error tracks that distance, the premise is load-bearing.","supporting_citations":[{"cited_title":"Bielitz, D","cited_arxiv_id":null,"evidence_quote":"Describes the autoregressive training framework (ATF) that SPF identifies as memory-intensive and uses as a comparison baseline."},{"cited_title":"Learning from Predictions: Fusing Training and Autoregressive Inference for Long-Term Spatiotemporal Forecasts","cited_arxiv_id":"2302.11101","evidence_quote":"Formulates learning from predictions for long-term spatiotemporal forecasting, setting the conceptual stage for training on the model's own outputs."},{"cited_title":"Cheng, M","cited_arxiv_id":null,"evidence_quote":"Supplies the shallow water benchmark used for the main ATF/PF/SPF comparisons."},{"cited_title":"Maulik, B","cited_arxiv_id":null,"evidence_quote":"Demonstrates the CAE-LSTM reduced-order architecture that SPF uses to compress solution fields before training."},{"cited_title":"Hochreiter, J","cited_arxiv_id":null,"evidence_quote":"Defines the LSTM surrogate model used as the dynamics predictor in all training frameworks."}],"review_version":2}