{"id":"4a29a9ab-31a1-4896-ad18-fa4e1bd45f52","arxiv_id":"2501.18122","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"VQLTI combines discrete latent intensity codes, FengWu forecast fields, and a potential-intensity feature to forecast tropical cyclone intensity, but the headline SOTA numbers rely on future reanalysis data as conditioning.","lead":"VQLTI is a machine learning system that forecasts tropical cyclone wind speed and pressure up to five days ahead by encoding current intensity into a learned discrete code and conditioning on future weather fields from the FengWu model. The reported 35 to 42 percent error reductions over ECMWF come from a setup that uses future reanalysis data as input, so the operational claim needs careful reading.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline SOTA claim rests on the non-real-time Table 1 protocol, whose conditioning inputs are not explicitly identified; if future ERA5 fields were used there, the claim is not operationally meaningful.","rationale":"The reader's weakest assumption is essentially the same concern: the headline claim's evaluation may condition on future ERA5 fields, making it non-apples-to-apples versus operational forecasts. My reading of the full text strengthens this. The 'Forecast' methodology section says accurate forecasting is possible 'if we obtain the discrete latent variables corresponding to the future TC intensity, as well as the future ERA5 data'. The long-term experiment section never restricts the decoder conditioning to FengWu forecast fields, while the real-time section explicitly states that future ERA5 is unavailable and therefore FengWu forecasts substitute for it. This asymmetry strongly implies Table 1 used future ERA5 conditioning, which would leak target information. The 'Base ERA5' row, at MSW errors of 16.86–18.52 m/s, shows raw future ERA5 near-surface wind is a poor direct proxy of IBTrACS MSW; a model conditioned on the full future 40x40x69 field patch (which contains geopotential, temperature, humidity, and winds at 13 levels in and around the storm) has ample information to infer the TC's intensity structure at that time, including its radial wind profile, pressure center, and thermodynamic state. VQLTI additionally receives the past intensity time series and the future PI computed from the same future fields, so the 4.30–5.05 m/s errors are plausible under oracle conditioning but say nothing about true predictive skill. The real-time 2022 experiment is the correct protocol and shows a much smaller, single-year advantage. The 2021 fine-tuning on the same year used for real-time evaluation also raises a mild data-leakage concern for the 2022 numbers, but the dominant concern remains the Table 1 protocol. The PI 'constraint' being a soft cross-attention feature rather than a hard bound is a secondary terminology issue, not the decisive flaw. Given the abstract explicitly advertises the 35.65%–42.51% reduction as the headline result, and the paper does not disclose the conditioning protocol in the Long-term Forecast section, the central claim is unverified as stated. The paper contains a legitimate engineering contribution and a plausible real-time evaluation, but the advertised SOTA claim needs re-statement and fair, significance-tested evaluation before it can be accepted. Verdict: REJECT, consistent with the reader.","tokens_in":13106,"tokens_out":2385,"duration_ms":20849,"concrete_test":"Re-run the 2019–2020 global evaluation replacing the future FengWu/ERA5 conditioning fields with fields available at forecast time (e.g., persisted latest ERA5 analysis or FengWu forecasts initialized at t=0, as in the real-time section), then recompute the 24–120h MSW MAE and the skill score versus ECMWF-IFS in Table 1. If the 35.65%–42.51% reduction collapses or shrinks to the single-year real-time margin, the headline claim as stated is not supported. Additionally, either the authors should state explicitly in Table 1's caption which conditioning inputs were used (ERA5 vs FengWu), or the paper should be revised to headline the 2022 real-time numbers.","verdict_should_be":"REJECT","load_bearing_attack":"The central advertised claim is the 35.65%–42.51% MSW error reduction versus ECMWF-IFS over 24–120h, presumably from Table 1. The paper never states in the Long-term Forecast section whether Table 1 uses future ERA5 fields as conditioning inputs. The Real-time Forecast section states 'future ERA5 fields are also unavailable, so we substitute them with FengWu's forecast results', implying the non-real-time protocol did use future ERA5 fields. Section 'Forecast' also notes 'If we obtain the discrete latent variables corresponding to the future TC intensity, as well as the future ERA5 data, we can then realize accurate forecasting', and the decoder is conditioned on the forecast field in the real-time pipeline only. If Table 1 decoded latent variables conditioned on future ERA5 spatial fields, then those fields carry information about the target cyclone's structure and intensity, including extreme wind values that are directly correlated with the MSW target. Conditioning on future states is not a forecast; it is interpolation with an oracle. The ECMWF-IFS and NCEP-GFS numbers, by contrast, are genuine forecasts made without future data. The 'Base ERA5' row shows 24h MSW MAE of 17.72 m/s (from simply reading off ERA5's maximum 10m wind speed in the future 10x10 degree box), while VQLTI reports 4.30. This enormous gap is consistent with the model learning to correct ERA5's known TC underestimation given the future box's full fields, not with skill that transfers to real-time. The real-time 2022 experiment is the only fair test, and there the claimed margin over NCEP-GFS shrinks dramatically (120h MSW: 5.53 vs 6.45, 14.26%) and is a single-year, single-run evaluation with no significance testing and baselines evaluated on different protocols.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VQLTI, a two-stage vector-quantized model for tropical cyclone intensity forecasting. Stage 1 uses a conditional VQ-VAE to map historical intensity into a discrete latent space conditioned on ERA5 spatial fields; stage 2 iteratively forecasts the latent variables with cross-attention and a sliding window, then decodes them into intensity using FengWu forecast fields as conditioning inputs and a potential-intensity (PI) physical constraint. The manuscript claims state-of-the-art 24-120 h global TC intensity forecasts, with MSW error reduced by 35.65%-42.51% relative to ECMWF-IFS, and additionally reports a 2022 real-time forecasting experiment and ablations over the VQ, FengWu-field, and PI components.","tokens_in":13374,"tokens_out":12050,"duration_ms":112443,"significance":"If the claimed results were obtained under a realistic forecast protocol, VQLTI would be a notable advance: it is among the first attempts to condition TC intensity forecasting on forecast fields from an ML weather model, and the discrete-latent formulation with a PI constraint is a plausible way to combat long-term error accumulation. The authors release code, and the ablation design cleanly separates the contributions of the codebook, FengWu fields, and PI. However, the headline quantitative claim is tied to a non-real-time evaluation that appears to condition on future ERA5 fields, which are unavailable at forecast time; the real-time experiment that avoids this is limited to one year, one run, and no significance testing. The significance of the work therefore rests on a protocol that the paper itself does not unambiguously define and that, as written, does not support the stated operational SOTA claim.","major_comments":[{"comment":"The abstract's headline 35.65%-42.51% MSW error reduction is tied to Table 1, but the Long-term Forecast section never states what conditioning inputs VQLTI receives. The Real-time Forecast section then says, 'Second, future ERA5 fields are also unavailable, so we substitute them with FengWu’s forecast results.' This reveals that the non-real-time evaluation (Table 1) conditioned the decoder on future ERA5 fields. Those fields contain the future atmospheric state of the storm, including wind structure directly correlated with the MSW target, so the comparison against ECMWF-IFS and NCEP-GFS—genuine forecasts made without future data—is not apples-to-apples. The Base ERA5 row in Table 1, with a 24 h MSW MAE of 17.72 m/s, shows how much signal the future box carries by itself. This is a load-bearing flaw for the SOTA claim; the authors must either rerun the 2019-2020 comparison using only inputs available at forecast time, or explicitly relabel Table 1 as an idealized oracle benchmark and remove it from the abstract and conclusion claims.","section":"Real-time Forecast; Table 1"},{"comment":"The real-time experiment is a single year (2022), a single model run, and reports no confidence intervals or significance tests, so the claimed real-time superiority is not statistically established. Moreover, VQLTI is fine-tuned on 2021 data, while the paper does not state that the DL baselines (TC-Pre, MSCAR, TCIF) receive any equivalent additional training or fine-tuning; if those baselines are evaluated with their original checkpoints trained only through 2017, the comparison gives VQLTI an extra-data advantage. The 35.85% reduction over ECMWF-IFS at 120 h in Table 2 therefore needs multi-year evidence and a matched-protocol description of baseline training before it can support an operational claim.","section":"Real-time Forecast; Table 2"},{"comment":"The input specification is internally inconsistent. Eq. (1) writes the model input as f(I_{-(n-1):0}, E_{-(n-1):0}, F_{1:m}, m), i.e., it always uses FengWu forecast fields F_{1:m} and no future ERA5, yet the Real-time Forecast section says that 'future ERA5 fields are also unavailable, so we substitute them with FengWu’s forecast results,' implying that earlier experiments used future ERA5. If Table 1 used future ERA5, Eq. (1) is incomplete; if it did not, the substitution sentence is misleading. The manuscript needs one unambiguous protocol statement that identifies, for each table and each row, exactly which future fields were used as conditioning inputs.","section":"Forecast; Eq. (1)"}],"minor_comments":[{"comment":"The sentence 'since FengWu does not perform forecasting' is confusing; the intended meaning is likely that FengWu does not forecast sea-surface temperature, which is why t2m is used as a substitute. Please rephrase.","section":"Potential Intensity"},{"comment":"Line 3 of Algorithm 1 sets 'score = CrossAttn(W[0], W[0], W[0])', which appears to hard-code the first window element rather than the current iteration index; as written, the attention computation over the window is not generally specified. Please clarify or correct the pseudo-code.","section":"Algorithm 1"},{"comment":"There are several typographical errors: 'ECWMF-IFS' should be 'ECMWF-IFS', 'JTCW' should be 'JTWC', and the Table 1 caption contains 'Western North Pacific Pacific'. Please correct these.","section":"Long-term Forecast; Table 1"},{"comment":"The text says a 10 degree longitude-latitude patch is cropped around the TC center, while Figure 2's caption mentions a '10 degree diameter'; please make the cropping geometry explicit and consistent.","section":"Data description; Figure 2"},{"comment":"The 'Base FengWu' row in Table 2 is not defined in the text; state whether it is computed exactly like Base ERA5 in Table 1 (e.g., max wind speed searched in the FengWu forecast box), since that affects the interpretation of VQLTI's improvement over FengWu.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"I weighed reject because the leading quantitative claim is invalidated by the oracle-conditioned protocol in Table 1. I ultimately recommend major_revision because the real-time experiment and the proposed method could support a more modest, defensible claim if the authors remove oracle-based comparisons from all headline statements, report a multi-year real-time evaluation with significance testing, and align the training protocols of the baselines. If multi-year real-time results cannot be produced, the authors should withdraw the SOTA claim and reframe the paper as a methods study."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the engineering is real and worth reading: the authors map TC intensity into a discrete latent space conditioned on spatial meteorological fields, then do multi-step forecasting in that space using FengWu forecast fields and a potential intensity (PI) feature. Second, the headline claim—a 35.65–42.51% MSW forecast error reduction over ECMWF-IFS for 24–120h—does not survive contact with the evaluation protocol. The main table appears to condition on future ERA5 fields, which makes it an oracle test, not a forecast.\n\nWhat is genuinely new: the specific combination of vector-quantized latent intensity, cross-attention conditioning on a learned weather model's forecast fields, and a PI-derived feature. The ablation study is honest and shows each component matters in the real-time setting. Code is released. Using FengWu forecast fields as a physical-knowledge input is a sensible direction and probably the paper's most useful idea.\n\nThe soft spots are not minor. The Long-term Forecast section never states whether Table 1 uses future ERA5 as conditioning. The Real-time Forecast section says \"future ERA5 fields are also unavailable, so we substitute them with FengWu's forecast results,\" which implies the non-real-time protocol did use future ERA5. If so, the comparison to NWP models is not apples-to-apples. The Base ERA5 row—17.72 m/s 24h MSW error from reading off future ERA5, versus VQLTI's 4.30—is consistent with the model learning to correct ERA5's known underestimation given the future box's full fields. The real-time experiment, the only fair test, shrinks the 120h margin over NCEP-GFS to 14.26%, is a single year and single run with no significance testing, and VQLTI is fine-tuned on 2021 while the baselines presumably are not. Also, the \"physical constraint\" from PI is a cross-attention feature, not a hard bound; the term overstates what the model does.\n\nWho is this for? Researchers building ML-based TC intensity forecasters. It deserves a serious referee: the real-time framing is a legitimate attempt, and the architectural choices are worth scrutiny. But the authors need to restate the headline around the real-time protocol, evaluate over multiple years with uncertainty quantification, and clearly label the oracle experiment as a diagnostic. The current abstract overclaims.","headline":"A promising integration of VQ latents, FengWu forecast fields, and potential intensity for TC intensity forecasting, but the headline SOTA claim is built on an oracle evaluation; the real-time evidence is too thin to support it as stated.","tokens_in":14014,"tokens_out":3778,"would_cite":false,"duration_ms":34306,"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":"This paper proposes VQLTI, a two-stage model that reframes tropical-cyclone intensity forecasting as a discrete-latent-space prediction problem and shows that injecting physical knowledge—the FengWu forecast field and a…","keywords":["tropical cyclone intensity","vector quantization","discrete latent space","potential intensity","FengWu forecast","physical constraints","long-term forecasting","ERA5"],"falsifier":"Run a strict re-forecast where every input is available at forecast time (no future ERA5 anywhere in the pipeline) and compare VQLTI's 120h MSW error to the reported real-time 5.53 m/s; if the error jumps well above that or fails to beat ECMWF-IFS on the same storms, the real-time claim is falsified. A second check is sensitivity: zero out the potential-intensity input and see whether the 120h forecast changes by more than training noise; if not, the physical constraint is not doing the claimed work.","tokens_in":12856,"feed_emoji":"🌀","tokens_out":9479,"duration_ms":84457,"temperature":0.7,"pith_summary":"Tropical-cyclone intensity forecasts out to five days are vital for disaster warning, but deep-learning models tend to accumulate error and lose skill beyond 24 hours. VQLTI's central idea is to stop forecasting raw intensity and instead forecast a sequence of discrete latent codes, where each code is learned to represent intensity together with its surrounding spatial weather pattern. The model then decodes those predicted codes into intensity, using the FengWu neural weather model's forecast field as a spatial condition and a potential-intensity (PI) value as a physical upper bound. On a global 2019–2020 test set, the authors report cutting maximum sustained wind (MSW) error by 35.65%–42.51% relative to ECMWF-IFS across 24–120h, and a real-time variant that substitutes FengWu forecasts for future reanalysis fields continues to beat existing deep-learning and operational baselines. If this holds, machine-learning models could move from short-term aids to credible five-day intensity guidance at a fraction of operational NWP cost.","feed_headline":"New model cuts 24-120h cyclone wind error by 35-43%","feed_subtitle":"VQLTI maps storm intensity to discrete codes constrained by physics and beats ECMWF-IFS on long-term forecasts.","key_machinery":"The engine is a two-stage conditional vector-quantized framework. Stage one is a CVQ-VAE: a Feature Pyramid Network extracts multi-scale feature maps from the 40×40 ERA5 patch, the encoder maps the intensity into a 1,024-entry discrete codebook via nearest-neighbor lookup, and the decoder reconstructs intensity conditioned on the same spatial features. Stage two is an iterative latent-space forecaster: a sliding window of past latent codes is passed through stacked cross-attention and an MLP to produce a future-code score, which is then cross-attended to the FengWu forecast field and to a potential-intensity value $V_{max, pmin} = PI(t2m, msl, t, q)$ (computed with Gilford's pyPI implementation) before being added residually to the last historical code. Each component carries a specific load: the codebook forces discrete, spatially-aware intensity representations; the FengWu field supplies the future atmospheric state in real-time use; and the PI term supplies a thermodynamic upper bound that counteracts the tendency of reanalysis and forecast fields to underestimate cyclone winds.","core_discovery":"The paper's central claim is that tropical cyclone intensity forecasting should be performed over a learned discrete latent space rather than over continuous intensity values, because the same intensity number can correspond to many different storm structures and environmental contexts. VQLTI first trains a conditional vector-quantized variational autoencoder (CVQ-VAE) that encodes the observed intensity into one of 1,024 learnable codebook entries while attending to multi-scale ERA5 spatial features, preserving both the intensity value and its spatial differences. Next, the model freezes the encoder and codebook, unfolds a sliding window of past latent codes, and iteratively predicts future codes with an attention-and-MLP module, using the FengWu forecast field and a computed potential-intensity value as cross-attention conditions and physical constraints. The authors report that this design slows error accumulation, yields state-of-the-art global 24–120h forecasts on 2019–2020 data, and that a fine-tuned real-time variant (with FengWu forecast fields substituted for future ERA5) remains superior to prior deep-learning methods, ECMWF-IFS, and NCEP-GFS through 120h.","pith_inferences":["Editorial inference: the headline 35.65%–42.51% improvement holds in a retrospective evaluation that conditions on future ERA5 fields; the real-time table shows a smaller but still positive advantage, so the operational gain is likely closer to the real-time numbers than to the headline.","Editorial inference: because the PI constraint and the FengWu conditioning are both derived from the same forecast field, the physical constraint may be partly redundant with the conditioning signal; using an independently observed sea-surface temperature to compute PI would test whether the constraint adds genuinely new information.","Editorial inference: real-time performance depends on how closely FengWu's forecast tracks the storm location, since long-lead forecasts drift spatially; a stress test with deliberately misplaced tracker positions would reveal the model's sensitivity to tracker errors.","Editorial inference: the discrete codebook maps intensities to a finite set of codes, which gives a natural categorical (e.g., Saffir-Simpson) reading; the same architecture could be extended to probabilistic forecasts by outputting a distribution over codebook entries."],"forward_implications":["If the 35.65%–42.51% MSW error reduction transfers to operations, a single learned model could replace or post-process NWP intensity forecasts, lowering the computing cost of early-warning systems.","The much slower error accumulation (48h MSLP error grows 16.01% versus 61–70% for prior deep-learning models) implies the latent-space iteration is the key to stable multi-day forecasts.","Because forecast length is a parameter and only the latent forecaster is trained, the same checkpoints can produce forecasts at arbitrary lead times and likely transfer to other meteorological target variables.","The real-time variant, though degraded relative to the retrospective setup, still beats all compared deep-learning baselines and ECMWF-IFS at 120h, suggesting the FengWu-substitution strategy is a workable path toward operational ML forecasts.","Ablation shows that codebook quantization, FengWu conditioning, and the PI constraint each contribute independently, so the architecture's components are modular and may generalize separately."],"supporting_citations":[{"why":"Supplies the vector-quantized time-series forecasting paradigm that the latent-space iteration of VQLTI extends.","marker":"Rasul et al. 2024"},{"why":"Provides the VQ-VAE codebook and commitment/codebook losses used in the first training stage.","marker":"Van Den Oord, Vinyals, et al. 2017"},{"why":"Defines the FengWu weather model and its 69-variable forecast fields, which VQLTI uses as conditioning and physical knowledge.","marker":"Chen et al. 2023a"},{"why":"Establishes the air-sea interaction theory yielding the potential-intensity upper bound used as a physics constraint.","marker":"Emanuel 1986"},{"why":"Provides the concrete PI computation (from t2m, msl, t, q) that the paper follows.","marker":"Gilford 2021"},{"why":"Supplies the IBTrACS best-track dataset giving the USA WIND and USA PRES ground-truth intensities.","marker":"Knapp et al. 2010"},{"why":"Supplies the ERA5 reanalysis fields used as training conditions and as spatial inputs in the retrospective evaluation.","marker":"Hersbach et al. 2020"},{"why":"The MSCAR baseline that VQLTI extends with vector quantization and one of the main deep-learning performance comparisons.","marker":"Wang et al. 2024"},{"why":"Defines the ECMWF-IFS and NCEP-GFS operational forecasts that serve as the primary NWP baselines in the SOTA comparison.","marker":"Bougeault et al. 2010"}],"fun_headline_variants":["VQLTI: physics-constrained AI slashes cyclone wind error 35-43%","Discrete latent codes keep cyclone forecasts sharp to 120h","AI storm model beats ECMWF-IFS by 35-43% on wind error","VQLTI reduces cyclone error accumulation with physical constraints","Physics-constrained VQLTI makes 120h cyclone intensity forecasts accurate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline accuracy numbers are measured with the model seeing the future atmospheric state as input, so the real-time claim depends on the assumption that a neural weather model's forecast fields can stand in for that future state without leaking the storm's actual intensity.","fun_headline_variants_meta":{"raw":{"variants":["VQLTI: physics-constrained AI slashes cyclone wind error 35-43%","Discrete latent codes keep cyclone forecasts sharp to 120h","AI storm model beats ECMWF-IFS by 35-43% on wind error","VQLTI reduces cyclone error accumulation with physical constraints","Physics-constrained VQLTI makes 120h cyclone intensity forecasts accurate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000744,"raw_usage":{"total_tokens":3342,"prompt_tokens":994,"completion_tokens":2348,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":2250}},"tokens_in":610,"tokens_out":2348,"duration_ms":16060,"temperature":1.0,"reasoning_tokens":2250,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T00:36:19.809707+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a strict re-forecast where every input is available at forecast time (no future ERA5 anywhere in the pipeline) and compare VQLTI's 120h MSW error to the reported real-time 5.53 m/s; if the error jumps well above that or fails to beat ECMWF-IFS on the same storms, the real-time claim is falsified. A second check is sensitivity: zero out the potential-intensity input and see whether the 120h forecast changes by more than training noise; if not, the physical constraint is not doing the claimed work.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the vector-quantized time-series forecasting paradigm that the latent-space iteration of VQLTI extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the air-sea interaction theory yielding the potential-intensity upper bound used as a physics constraint."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the concrete PI computation (from t2m, msl, t, q) that the paper follows."},{"cited_title":"R.; Kruk, M","cited_arxiv_id":null,"evidence_quote":"Supplies the IBTrACS best-track dataset giving the USA WIND and USA PRES ground-truth intensities."},{"cited_title":"H.; Ebert, B.; Fuentes, M.; Hamill, T","cited_arxiv_id":null,"evidence_quote":"Defines the ECMWF-IFS and NCEP-GFS operational forecasts that serve as the primary NWP baselines in the SOTA comparison."}],"review_version":1}