{"id":"79ed7024-28e5-43cc-a02c-f37f670bf2b0","arxiv_id":"2506.08049","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"TelePiT claims state-of-the-art global S2S forecast skill on ERA5 with a transformer that uses zonal-mean inputs, a multi-scale physics-inspired ODE, and teleconnection-biased attention, but key verification evidence is missing.","lead":"A new deep learning model for 3-to-6-week weather forecasts combines zonal-average embeddings, a physics-inspired neural ODE, and teleconnection-style attention, reporting large gains over published ML and operational baselines. The paper omits a key published S2S model, releases no code, and its architecture discards longitude information, so the headline results are not yet verifiable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The model's input is a per-latitude zonal mean (Sec. 3.1, Eqs. 2-5), so its forecasts are invariant to all longitude-resolved initial structure; this leaves no mechanism for the claimed MJO/teleconnection awareness and makes the reported full-field S2S superiority unsubstantiated.","rationale":"The reader's weakest_assumption identifies the same bottleneck, and I agree. The zonal-mean input is not a stylistic simplification; it is a hard information-theoretic filter. Because every module downstream consumes only tokens that are longitude-averaged, the model's output is invariant under any longitudinal rearrangement of the initial state. Yet the S2S predictability the paper claims—MJO, NAO, extratropical teleconnections, regional t2m anomalies—is precisely longitude-resolved. The full-field metrics in Tables 1-3 may still look good if the targets are dominated by zonal-mean variance and seasonal cycle, but ACC is anomaly-based and still extremely high, which makes the numbers difficult to reconcile with the architecture. The proposed concrete test would separate zonal-mean skill from longitude-resolved skill. I would keep the reader's REJECT: as written, the central claim is not supported. I note the paper also lacks code and error bars and omits FuXi-S2S, but those are secondary; the architectural information loss is the load-bearing issue.","tokens_in":44444,"tokens_out":12701,"duration_ms":148656,"concrete_test":"Compute longitude-anomaly skill on the 2018 test set: for each target field, subtract the per-latitude zonal mean from both prediction and target, and report RMSE and ACC on these de-meaned fields. Compare TelePiT against (a) calendar-day climatology and (b) a simple zonal-mean-only regression model that predicts each grid point from the zonal-mean input state. If TelePiT does not significantly beat these null models on the longitude-anomaly component, the reported full-field metrics are dominated by zonal-mean/climatological predictability and the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing weakness is the input information bottleneck. In Sec. 3.1, each token is built from u_i = (1/W)Σ_j X_{i,j}, and the longitude embedding is averaged over W before being added to the latitude embedding (Eqs. 4-5). Consequently the entire stack (multi-scale ODE and teleconnection transformer) sees only the zonal-mean state at t1; no longitude-resolved information and no date or temporal history reaches the model. The output head (Eq. 19) can emit arbitrary per-longitude values, but those values are a fixed function of latitude and the zonal-mean input, so the model cannot condition on the phase of any zonally asymmetric pattern. This is internally inconsistent with the paper's central narrative: the MJO (Sec. 3.3.1) is defined by its zonal phase structure, and teleconnections such as the NAO and Rossby wave trains are longitude-specific. Unless the reported RMSE/ACC/SpecDiv gains are entirely dominated by the zonal-mean/climatological component of the targets, the architecture has no mechanism to produce them. The ablation (Table 4) highlights the problem: the largest claimed gain is for t2m (28.5 to 12.1 K), a field whose predictable anomalies are strongly tied to regional land-surface and SST patterns, none of which survive zonal averaging. This is not merely a missing baseline; it is a structural mismatch between the claimed mechanism and the actual computation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"TelePiT is a deep-learning architecture for global subseasonal-to-seasonal (S2S) forecasting from ERA5 reanalysis data. The model encodes the initial state as per-latitude zonal means, evolves them through a multi-scale physics-informed neural ODE and a teleconnection-aware transformer, and outputs full longitude-resolved forecasts for weeks 3-4 and 5-6. The paper claims that this lightweight model outperforms both data-driven baselines and operational numerical weather prediction systems across all forecast horizons, with the largest gains for near-surface variables. The evaluation uses a ChaosBench-based protocol with RMSE, ACC, SpecDiv, MS-SSIM, and SpecRes, and includes an ablation study and an out-of-sample 2019 robustness check.","tokens_in":44802,"tokens_out":6125,"duration_ms":68814,"significance":"If the central claim held, this would be a striking result: a model that sees only zonal averages at a single initial time would produce the most skillful global S2S forecasts available, at a fraction of the computational cost of existing models. The paper deserves credit for its broad multi-metric evaluation, the explicit 2019 out-of-sample test, and its honest acknowledgment in Appendix A.3 that the latent ODE is inspired by, rather than derived from, atmospheric equations. However, the architecture irreversibly removes longitude-resolved information from the input, which is inconsistent with the paper's central narrative of teleconnection awareness and with the claimed full-field forecast skill. The omission of FuXi-S2S, a cited state-of-the-art baseline, further weakens the comparative claim. As a result, the findings are not currently established.","major_comments":[{"comment":"The model's only input is the zonal mean per latitude, u_i = (1/W) sum_j X_{i,j}, and the longitude embedding is averaged before being added to the latitude embedding (Eqs. 4-5). The output head then maps each latitude token to all longitudes via Eq. (19). Consequently, the forecast is a function of latitude and the zonal-mean state only; it is invariant under arbitrary longitude shifts of the initial condition. This removes any mechanism for the teleconnection awareness claimed in Section 3.3.1: the MJO, NAO, and Rossby wave trains are defined by their zonal phase structure, which the architecture cannot see. A concrete falsifiable test is to shift the initial field by a longitude offset and confirm that the forecast does not change; the paper should also compare against a per-longitude climatology or persistence baseline. Until such a test is provided, the RMSE/ACC/SpecDiv gains in Tables 1-4 could be entirely dominated by the zonal-mean/climatological component of the targets.","section":"Section 3.1, Eqs. (2)-(5); Section 3.4, Eq. (19)"},{"comment":"The paper cites FuXi-S2S (ref [8]) as a strong global S2S model but does not include it in the comparison tables. The abstract and Section 4.1 claim that TelePiT 'significantly outperforms state-of-the-art data-driven baselines' and 'across all forecast horizons.' Without FuXi-S2S, which is a directly relevant published S2S model, the state-of-the-art claim is not supported by the evidence presented.","section":"Table 1 and Section 4.1"},{"comment":"The near-surface metrics are implausibly large in magnitude: u10 RMSE for TelePiT is 0.491 m/s versus 4.0 m/s for FourCastNetV2 and PanguWeather, and t2m RMSE is 12.1 K versus 60.9 K for ClimaX. For a model that ingests only zonal means, these gaps are suspicious and suggest either different evaluation conventions (for example, anomaly RMSE versus absolute RMSE, or different weighting/gridding) or a trivial solution such as predicting climatology. The paper should state explicitly whether these are absolute or anomaly metrics, and should add a climatology baseline to the comparison.","section":"Table 1, Section 4.1.1"},{"comment":"The theoretical claims that the learned teleconnection patterns converge to a basis spanning the EOF subspace (Prop. A.8) and that the frequency bands preserve (1-delta) of the mutual information with the target (Prop. A.4) are asserted with proof sketches that assume the conclusion. Specifically, Prop. A.8's proof says the optimization 'is equivalent' to variance maximization without establishing that equivalence, and Prop. A.4 does not prove the existence of a band decomposition satisfying the stated mutual-information lower bound. These propositions are presented as theoretical support but do not provide it; they should be corrected or removed.","section":"Appendix A.4, Proposition A.8 (Eq. 31); Appendix A.2, Proposition A.4 (Eq. 25)"}],"minor_comments":[{"comment":"The header reads 'SepcDiv' instead of 'SpecDiv'.","section":"Table 2"},{"comment":"The ACC formula does not include the latitude weight w(theta_i) in the numerator, while the denominator includes it. Please verify whether this is a typo or an intentional definition.","section":"Appendix B.1, Eq. (36)"},{"comment":"The table contains '9404' and '0556' where '0.9404' and '0.0556' are intended; there are also inconsistencies between 'SpecRec' and 'SpecRes' in Tables 7 and 8.","section":"Table 6"},{"comment":"The term 'learnable wavelet decomposition' is used, but Eq. (9) is an MLP followed by a split; no wavelet properties such as orthogonality or multiresolution analysis are enforced. Clarify the terminology.","section":"Section 3.2.1 and Eq. (9)"},{"comment":"Section 3.2.2 states zero-padding boundary conditions (x_0 = x_{H+1} = 0), while Appendix A.3 states 'we apply periodic boundary conditions.' One of these descriptions is incorrect.","section":"Section 3.2.2 vs. Appendix A.3"}],"recommendation":"reject","confidential_remarks":"The structural bottleneck identified in Section 3.1 is central and, on the evidence of the manuscript, cannot be fixed within the current scope: the model would need a redesign to use longitude-resolved input, and the entire experimental evaluation would need to be rerun. The missing FuXi-S2S baseline and the implausible near-surface metrics reinforce the need for rejection rather than major revision. I would not recommend the editor invite a revision unless the authors are willing to make the architectural change and re-run the experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is not junk. The architecture is clearly described, the appendix is unusually thorough, and the authors do real work: a multi-scale physics-informed ODE, an attention bias for global patterns, and an out-of-sample 2019 robustness test. If the zonal-mean trick worked, it would be a useful baseline for the S2S community. But the central claim does not survive contact with the model's own equations.\n\nThe soft spot is load-bearing, not cosmetic. In Section 3.1, each token is the zonal mean u_i, and the longitude embedding is averaged out before the stack. The transformer and ODE therefore never see any longitude-resolved information. The output head can emit per-longitude values, but those values are a fixed function of latitude and the zonal-mean initial state. There is no mechanism to condition on the phase of the MJO, a Rossby wave train, or any other zonally asymmetric pattern. Yet the abstract and Section 3.3.1 claim explicit teleconnection modeling. That is not a missing baseline; it is a structural mismatch between narrative and computation. The stress-test note is right.\n\nOther problems are proportional. FuXi-S2S, a strong published S2S model that the paper cites, is absent from the comparison tables. The near-surface numbers are hard to believe: u10 RMSE of 0.49 m/s versus 4.0 for Pangu and FourCastNetV2 suggests the iterative baselines are not being applied in the regime they were built for, and the directly trained ClimaX/CirT may not be fairly tuned. There are no error bars and no code. Proposition A.6's Lipschitz bound is simply wrong: the tanh does not make the whole right-hand side γ-Lipschitz; the diffusion, advection, and MLP terms have their own Lipschitz constants. Propositions A.4 and A.8 are partly circular, assuming the very preservation and EOF-convergence they claim to prove.\n\nThat said, the empirical tables are internally consistent, the ablation study is informative, and the 2019 robustness check at least shows the model does not immediately fall apart on a new year. The paper is a coherent engineering exercise; it is just not the scientific advance the abstract advertises.\n\nWho is this for? Someone working on S2S forecasting with ML who wants to see a creative but ultimately flawed attempt, and who cares about evaluation hygiene. It could also serve as a teaching example of why you must trace what information actually enters a model. I would send it to a careful referee who knows atmospheric dynamics, with a clear request to check the input bottleneck and the baseline protocol. It deserves peer review, but the likely outcome after revision is a much more modest paper about a zonal-mean benchmark, not a state-of-the-art forecast system.","headline":"A clearly written architecture paper whose headline skill claim collapses on the input bottleneck: the model only sees zonal means, so the teleconnection awareness it advertises cannot exist; still worth a referee's time as a cautionary case.","tokens_in":45318,"tokens_out":2025,"would_cite":false,"duration_ms":27109,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","86A10","62M20"],"pacs":[],"model":"deepseek-v4-flash","headline":"TelePiT claims that a model reading only latitude-averaged atmosphere, evolved by a physics-informed ODE with teleconnection-aware attention, produces the most skillful global 3-to-6-week forecasts yet reported, cutting 2-meter…","keywords":["subseasonal-to-seasonal forecasting","global climate prediction","physics-informed neural ODE","teleconnection-aware attention","spherical harmonic embedding","multi-scale wavelet decomposition","deep learning weather forecasting","ERA5 reanalysis"],"falsifier":"Collect pairs of initialization dates whose latitude-averaged fields are nearly identical but whose longitude-resolved states differ sharply, for example opposite Madden–Julian Oscillation phases. TelePiT will issue nearly identical forecasts for the two members of every such pair, because longitude information never enters the model. If the verifying weather then diverges between the paired cases, as MJO theory predicts for winter temperatures and precipitation over North America, the model is provably blind to a predictable, physically important S2S signal.","tokens_in":44241,"feed_emoji":"🌍","tokens_out":16773,"duration_ms":163815,"temperature":0.7,"pith_summary":"This paper claims that subseasonal-to-seasonal forecasts — climate outlooks three to six weeks ahead — can be made more accurately by a model that reads only the average atmospheric state around each latitude circle, discarding all longitude structure. The proposed architecture, TelePiT, feeds these zonal means through a learned multi-band decomposition, evolves each band with a physics-inspired differential equation containing advection, diffusion, forcing, and a neural correction, and applies a transformer whose attention is biased by learned teleconnection patterns. Evaluated on the ERA5 reanalysis under the ChaosBench protocol, it reports lower error than every data-driven baseline and every operational numerical weather prediction system across all tested variables and both forecast horizons, with 2-meter temperature RMSE cut by over 57% relative to the previous best model and spectral divergence roughly an order of magnitude smaller. If these results hold, a lightweight model running in milliseconds would set the standard for global subseasonal-to-seasonal skill.","feed_headline":"Zonal-mean inputs alone beat top 6-week forecast systems","feed_subtitle":"TelePiT cuts 2-meter temperature error by over 57% and beats operational models at a fraction of the compute.","key_machinery":"The argument rides on three interlocking mechanisms. First, the Spherical Harmonic Embedding reduces each day's 63-variable globe to a sequence of 121 latitude tokens by zonal averaging — formally a projection onto the $m=0$ spherical harmonics — destroying all longitude structure while keeping the latitude profile. Second, a learnable wavelet decomposition splits these tokens into $L+1$ frequency bands, each evolved by a physics-informed neural ODE whose right-hand side is $\\gamma\\,\\tanh\\!\\big(\\boldsymbol{\\nu}\\odot(\\mathbf{x}_{i+1}-2\\mathbf{x}_i+\\mathbf{x}_{i-1}) + \\boldsymbol{\\mu}\\odot(\\mathbf{x}_{i+1}-\\mathbf{x}_{i-1})/2 + \\mathbf{f} + \\alpha\\,\\mathrm{MLP}(\\mathbf{x}_i)\\big)$, i.e. explicit diffusion, advection, forcing, and a learned correction. Third, the Teleconnection-Aware Transformer adds a bias to the attention logits: a learned combination of teleconnection pattern vectors $\\mathbf{P}_j$ forms a query whose dot products with all latitude keys are added to the standard attention scores, letting one latitude band attend to others according to the active global climate mode. A cross-scale fusion layer then mixes all frequency bands before a shared decoder emits full longitude-resolved fields for both target windows.","core_discovery":"TelePiT's central claim is that explicitly structuring a forecasting model around three physical ideas — spherical geometry, scale-separated atmospheric dynamics, and long-range climate couplings — produces global S2S forecasts more accurate than any existing data-driven or operational system at both weeks 3–4 and weeks 5–6. Concretely, the paper reports z500 RMSE of 48.67 gpm at weeks 3–4 versus 53.81 gpm for CirT and 63.67 gpm for ECMWF; 2-meter temperature RMSE of 12.06 K versus 28.53 K for CirT and 60.89 K for ClimaX, with ACC of 0.996; and SpecDiv values of 0.0018–0.018 across key variables, one to two orders of magnitude below the best baselines. Ablation results attribute the gains to all three components, with the spherical harmonic embedding most critical for near-surface temperature, the wavelet decomposition most important for geopotential height, and the teleconnection bias increasingly valuable at the longer lead time. The authors interpret the results as evidence that a single model can capture the multi-scale physical processes and teleconnections that govern predictability beyond the two-week deterministic limit.","pith_inferences":["My reading: the measured gains may come chiefly from a strong smoothing prior — the zonal-mean tokenization plus the transport-form ODE suppresses high-wavenumber noise — rather than from the named mechanisms; the paper itself concedes in its appendix that the ODE is 'inspired by, rather than derived from' atmospheric equations, and the teleconnection patterns are simply trainable vectors never ch","Testable extension: replacing the single zonal mean per latitude with a few longitude sectors would directly measure how much remaining predictable signal lives in the east–west structure the model currently erases, and would show whether MJO-phase information adds or subtracts skill.","If the zonal-mean bet is right, S2S predictability research could shift toward the axisymmetric state — polar-vortex strength, meridional temperature gradients, tropical heating asymmetries — as the natural input variables for weeks 3–6, with the full three-dimensional field as the decode target.","Because 2-meter temperature in ERA5 is itself a model-derived field, the headline t2m numbers (RMSE 12.06 K, ACC 0.996) deserve an independent check against station-based or otherwise observation-grounded temperature analyses before being read as skill against the real near-surface atmosphere."],"forward_implications":["If the reported skill is real, operational S2S outlooks could be issued at a fraction of current cost: TelePiT runs at 14.5 GFLOPs with 37 million parameters, roughly 7,500 times fewer FLOPs than GraphCast and 11,500 times fewer than Pangu-Weather, with inference on the order of tens of milliseconds.","Each architectural component earns its place according to the ablations: removing the spherical harmonic embedding raises t2m RMSE from 12.1 K to 27.2 K, removing the wavelet decomposition costs 7–8% on geopotential height, and removing the teleconnection bias hurts most at weeks 5–6.","Spectral fidelity is the paper's largest margin: SpecDiv values of 0.0018–0.018 across variables are one to two orders of magnitude below every baseline, implying forecast power spectra much closer to the verification fields.","The advantage persists out-of-sample on the 2019 holdout year, which the paper takes as evidence that the model captures generalizable atmospheric structure rather than training-period patterns.","A single instantaneous zonal-mean snapshot suffices for both target windows at once, a direct corollary of the design that makes the reported skill a statement about how much predictable S2S signal sits in the latitude profile alone."],"supporting_citations":[{"why":"Supplies the ERA5 reanalysis training, validation, and test fields (1979–2018) that every forecast in the paper is trained and verified on.","marker":"[15]"},{"why":"The ChaosBench benchmark provides the dataset processing, the SpecDiv spectral-fidelity metric, and the operational-model scores (ECMWF, UKMO, NCEP, CMA) that TelePiT is compared against.","marker":"[32]"},{"why":"CirT is the geometry-inspired transformer baseline that TelePiT most directly extends and the strongest data-driven rival it must beat.","marker":"[25]"},{"why":"Pangu-Weather is the pretrained data-driven baseline that TelePiT outperforms at all reported variables and horizons.","marker":"[6]"},{"why":"FourCastNetV2 is the pretrained data-driven baseline whose RMSE, ACC, and SpecDiv numbers anchor the model comparison.","marker":"[23]"},{"why":"GraphCast is the pretrained baseline that anchors the medium-range-to-S2S comparison at weeks 3–4.","marker":"[24]"},{"why":"ClimaX is the foundation-model baseline retrained under the same configuration as TelePiT for a fair comparison.","marker":"[33]"},{"why":"ClimODE supplies the latent-space neural ODE formulation that the physics-informed ODE component explicitly draws from.","marker":"[44]"},{"why":"The Kullback–Leibler divergence principles underly SpecDiv, the metric on which TelePiT shows its largest margin.","marker":"[22]"},{"why":"FuXi-S2S establishes the global S2S task formulation — predicting weeks 3–4 and 5–6 means from a single initial state — that TelePiT follows.","marker":"[8]"}],"fun_headline_variants":["TelePiT beats ECMWF and top AI at 3-6 week forecasts","Physics-informed transformer cuts S2S temp error by 57%","AI model with teleconnection bias tops operational S2S systems","Multi-scale physics in transformer improves subseasonal skills"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model sees only the average state around each latitude circle, so any forecast signal that lives in the east–west arrangement of the atmosphere — such as the traveling tropical rain clusters of the Madden–Julian Oscillation — is erased before training begins, and both the reported skill and the teleconnection story rest on that information being unnecessary.","fun_headline_variants_meta":{"raw":{"variants":["TelePiT beats ECMWF and top AI at 3-6 week forecasts","Physics-informed transformer cuts S2S temp error by 57%","AI model with teleconnection bias tops operational S2S systems","Multi-scale physics in transformer improves subseasonal skills"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1501,"prompt_tokens":1025,"completion_tokens":476,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":402}},"tokens_in":641,"tokens_out":476,"duration_ms":6349,"temperature":1.0,"reasoning_tokens":402,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:41:35.349729+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect pairs of initialization dates whose latitude-averaged fields are nearly identical but whose longitude-resolved states differ sharply, for example opposite Madden–Julian Oscillation phases. TelePiT will issue nearly identical forecasts for the two members of every such pair, because longitude information never enters the model. If the verifying weather then diverges between the paired cases, as MJO theory predicts for winter temperatures and precipitation over North America, the model is provably blind to a predictable, physically important S2S signal.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The ChaosBench benchmark provides the dataset processing, the SpecDiv spectral-fidelity metric, and the operational-model scores (ECMWF, UKMO, NCEP, CMA) that TelePiT is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CirT is the geometry-inspired transformer baseline that TelePiT most directly extends and the strongest data-driven rival it must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Pangu-Weather is the pretrained data-driven baseline that TelePiT outperforms at all reported variables and horizons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FourCastNetV2 is the pretrained data-driven baseline whose RMSE, ACC, and SpecDiv numbers anchor the model comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GraphCast is the pretrained baseline that anchors the medium-range-to-S2S comparison at weeks 3–4."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ClimODE supplies the latent-space neural ODE formulation that the physics-informed ODE component explicitly draws from."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FuXi-S2S establishes the global S2S task formulation — predicting weeks 3–4 and 5–6 means from a single initial state — that TelePiT follows."}],"review_version":1}