Pith. sign in

REVIEW 3 major objections 37 references

Routing only a quarter of weather-grid tokens through expensive middle transformer layers, then scattering residual updates, improves forecasts while cutting training cost roughly 2.5×.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-12 06:45 UTC pith:QEEE76XJ

load-bearing objection Clean residual sparse routing that actually improves ERA5 skill while cutting ViT training cost ~2.5×; single-run tables are the main soft spot, not a load-bearing collapse. the 3 major comments →

arxiv 2607.02829 v1 pith:QEEE76XJ submitted 2026-07-02 cs.LG

Less Tokens, Better Forecasts: Sparse Residual Routing for Efficient Weather Prediction

Janet Wang , Yunbei Zhang , Lin Zhao , Xi Xiao , Jihun Hamm , Xiao Wang This is my paper
classification cs.LG
keywords weather forecastvision transformerefficient trainingsparse token routingresidual deltaERA5diffusion models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

ViT-based weather models spend the same compute on every spatial token even though large regions of the atmosphere are smooth and redundant. Standard pruning or merging drops or fuses tokens, which breaks dense fixed-grid forecasting and autoregressive rollout. This paper introduces Sparse-Reslim, a parameter-free plug-in: it sends only 25% of tokens through the middle blocks, treats those blocks as residual updates, and scatters only the deltas back so unselected tokens keep their exact pre-routing states. Across ERA5 resolutions up to 0.25° and both a deterministic Transformer and a diffusion model, the method raises accuracy on every evaluated variable while speeding training about 2.5× (up to 3.18× at 0.25°) and cutting peak memory more than 2.2×. A controlled breakdown attributes most of the accuracy gain to sparse residual routing itself, with random selection adding a free regularization benefit.

Core claim

Sparse residual routing is compatible with dense spatiotemporal weather prediction: processing only a random 25% of spatial tokens in the middle transformer blocks, computing the residual change for those tokens, and scattering that delta back leaves every grid cell with a physically meaningful representation, and this design both accelerates training and improves forecast skill over the dense baseline on every evaluated variable and both model families.

What carries the argument

Sparse-Reslim: a three-stage dense–sparse–dense block schedule that gathers a random keep-ratio subset, runs middle blocks only on that subset, and scatters residual deltas (processed minus pre-routing) so unselected tokens retain identity paths with no mask embeddings, fusion layers, or extra parameters.

Load-bearing premise

Randomly keeping a fixed fraction of spatial tokens and only updating those with residual deltas still carries enough intermediate atmospheric information for accurate dense-grid forecasts and stable multi-day rollout, without systematic geographic bias.

What would settle it

Train the identical dense and Sparse-Reslim models at the same resolution and lead time; if Sparse-Reslim at r=0.25 fails to beat the dense baseline on Z500, T850, T2m, and U10 RMSE/ACC, or if 10-day autoregressive skill collapses relative to dense, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Existing ViT weather forecasters can enable the routing flag and train ~2.5× faster with lower peak memory without redesigning the backbone.
  • Accuracy gains appear across both deterministic and generative (diffusion) pipelines, so the same residual-routing pattern applies to ensemble-style models.
  • Benefits strengthen at higher resolution and larger model depth, so operational 0.25° systems stand to gain the most.
  • Sparse routing stacks with mixed precision and resolution choices rather than replacing them.
  • Autoregressive multi-day rollout remains stable because unselected tokens keep true pre-routing states rather than masks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same residual-delta identity path could transfer to other dense fixed-grid geophysical tasks (ocean, land-surface, or air-quality) where every cell must stay valid.
  • If adaptive selection later beats random at storm-scale resolutions, the residual scatter layer can stay fixed while only the selector changes.
  • Training-time sparsity may leave inference-time ensemble averaging over random masks as a cheap diversity source without retraining.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 0 minor

Summary. The paper proposes Sparse-Reslim, a parameter-free plug-in for ViT-based weather models that partitions transformer blocks into dense early, sparse middle, and dense late stages. At the sparse stage it randomly keeps r=0.25 of spatial tokens, runs only those through the middle blocks, and scatters residual deltas back so unselected tokens retain their exact pre-routing representations. This is intended to cut quadratic attention cost while preserving a full fixed-grid state for dense prediction and autoregressive rollout. Experiments on ERA5 (1.40625°, 1.0°, 0.25°) with a deterministic Res-Slim-ViT and an EDM diffusion backbone report ~2.5× training speedup (up to ~3.18× at 0.25°), >2.2× peak-memory reduction, and improved RMSE/ACC/SSIM on T2m, U10, Z500, and T850 relative to the dense baseline, with ablations on keep ratio, block split, selection strategy, stochastic depth, and a pure-ViT backbone.

Significance. If the accuracy-plus-efficiency result holds under replication, Sparse-Reslim is a practical contribution for a field where ViT training cost is a primary bottleneck at operational resolution. The residual-delta formulation is a clean, parameter-free adaptation of dense–sparse–dense routing to fixed-grid spatiotemporal prediction, avoiding mask embeddings and fusion layers used in image-generation methods such as Sprint. Strengths include multi-resolution ERA5 evaluation, transfer to both deterministic and generative backbones, explicit efficiency measurements (GFLOPs, wall-clock, VRAM), and controlled ablations that separate sparsity from stochastic regularization. Code is stated to be available. These elements make the work useful even if some gains prove partly regularization-driven.

major comments (3)
  1. Table 1, Fig. 4, and the pure-ViT results (Table 11) report single-run point estimates for every variable and resolution. Sparse-Reslim draws a fresh random permutation every forward pass (Sec. 3.2; Algorithms 1–2), so training is stochastic; the paper never reports multi-seed means, standard deviations, or significance tests. The central claim that accuracy improves on every evaluated variable is therefore not yet statistically established, especially for smaller absolute gains (e.g., U10 at 1.40625°). At least 3 seeds (or bootstrap intervals) on the main 1.0° deterministic setting are needed to support that half of the claim.
  2. The abstract and introduction emphasize compatibility with autoregressive rollout, yet Table 3 reports only T2m ACC for a 1–10 day rollout at 1.40625° after 24 h training. Main quality tables (Table 1, Fig. 4) are single-step 120 h forecasts. Load-bearing evidence for the residual-identity path under multi-step error accumulation is therefore thin; at least Z500/T850 multi-day rollouts (and preferably the generative model) should be shown before the rollout-stability claim can be taken as established.
  3. Sec. 3.2 and Table 5 argue that random selection is preferable to attention-score or learned selection because it avoids geographic bias and adds free regularization. The paper does not quantify whether random keep-ratio sampling systematically under-samples dynamically active regions (jet streams, fronts) over long training, nor does it report spatial error maps or regional skill scores. A short diagnostic (e.g., selection frequency vs. latitude/variance, or regional RMSE) would test the weakest modeling assumption that a fixed random subset preserves intermediate atmospheric representations adequately.

Circularity Check

0 steps flagged

No significant circularity: empirical method and ERA5/dense-baseline results are self-contained; only minor non-load-bearing self-citation of the ORBIT-2 backbone architecture.

full rationale

The paper proposes Sparse-Reslim (dense-early / sparse-middle residual-delta / dense-late routing with random keep-ratio r=0.25) as a parameter-free plug-in and evaluates it experimentally against the identical dense baseline (r=1) on held-out ERA5 years, reporting RMSE/ACC/SSIM and wall-clock/memory metrics. The residual-delta construction (Eq. 9 / Algorithms 1-2) is defined so that unselected tokens retain pre-routing representations exactly; this is an engineering choice that preserves the fixed grid by construction, but the accuracy and speedup claims are not derived from that definition—they are measured outcomes. Keep-ratio and block-split choices are ablated (Fig. 5, Table 4) rather than fitted to the target skill scores and then re-presented as predictions. Random selection is justified by zero-overhead regularization, not by a uniqueness theorem. The sole self-citation of note is the Res-Slim-ViT backbone adapted from ORBIT-2 [28] (shared author Xiao Wang); it supplies only the dual-path architecture used for one of the two model families and is not invoked to force the sparse-routing result or to forbid alternatives. Pure-ViT and EDM experiments further isolate the contribution. No self-definitional loop, no fitted-input-as-prediction, no load-bearing uniqueness import, and no renaming of a known result appear. The derivation chain is therefore ordinary empirical ML work resting on external data splits and controlled ablations.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The central claim is empirical and engineering-driven. It rests on standard ViT/attention cost assumptions, the domain premise that intermediate atmospheric tokens are redundant, and a small set of hand-chosen routing hyperparameters (keep ratio, block split) that are ablated but still free design choices. No new physical entities are postulated.

free parameters (3)
  • keep_ratio r = 0.25
    Fraction of tokens routed through middle blocks; default 0.25 chosen by ablation (Fig. 5) as accuracy-speed sweet spot.
  • block_split (n1,n2,n3) = (2,8,2) for N=12
    Partition of transformer depth into dense-early / sparse-middle / dense-late; default (2,8,2) for N=12 chosen by Table 4 ablation.
  • patch_size at 0.25° = 8
    Set to 8 so routed token count matches the 1.0° setting rather than reflecting native high-resolution density.
axioms (4)
  • standard math Self-attention cost is O(L²D) per block and dominates middle-layer compute for large L.
    Used in Sec. 3.2 computational-savings derivation and theoretical speedup formula.
  • domain assumption Atmospheric fields are spatially autocorrelated so intermediate middle-layer representations can be approximated by a random subset plus residual scatter.
    Stated in Introduction and Sec. 3.2 motivation; load-bearing for why sparse routing does not destroy dense prediction.
  • ad hoc to paper Unselected tokens that retain exact pre-routing representations remain physically meaningful for fixed-grid autoregressive rollout.
    Core residual-delta design choice (Eq. 9, Algorithms 1–2); validated only by the paper’s own rollout and skill tables.
  • domain assumption ERA5 reanalysis (1979–2017 train / 2018–19 val / 2020 test) is a valid proxy for evaluating medium-range forecast skill.
    Standard community protocol (ClimaX/WeatherBench lineage) adopted in Sec. 4.1.
invented entities (1)
  • Sparse-Reslim residual-delta routing module no independent evidence
    purpose: Make sparse middle-block processing compatible with dense fixed-grid weather prediction without mask tokens or fusion parameters.
    New architectural component introduced in Sec. 3.2; independent evidence is the empirical speed/accuracy tables rather than an external physical prediction.

pith-pipeline@v1.1.0-grok45 · 23890 in / 2995 out tokens · 24809 ms · 2026-07-12T06:45:56.664357+00:00 · methodology

0 comments
read the original abstract

Existing ViT-based weather forecasting models apply uniform computation across all spatial tokens, even though nearby atmospheric grid points often contain similar values and large regions evolve smoothly over time. This makes much of the intermediate per-token computation redundant. Standard token-efficiency methods, such as pruning or merging, reduce cost by removing or fusing tokens. However, weather forecasting is a spatiotemporal dense prediction problem in which a history of atmospheric states must be mapped to future values on the original latitude-longitude grid. Thus, every grid cell must retain a physically meaningful representation, especially under autoregressive rollout. We introduce Sparse-Reslim, a parameter-free plug-in routing module that makes sparse token processing compatible with this fixed-grid requirement. Sparse-Reslim routes only 25% of spatial tokens through the expensive middle transformer blocks and treats those blocks as residual updates: it computes the change produced for the routed tokens and scatters only this delta back to the full sequence. Unselected tokens keep their pre-routing representations exactly, so no grid cell is dropped or replaced by a mask token, and no fusion layer or additional parameters are introduced. Across ERA5 resolutions up to the operational 0.25\textdegree{} standard and two model families, a deterministic Transformer and a diffusion model, Sparse-Reslim improves forecast accuracy on every evaluated variable while substantially reducing cost: training is about 2.5x faster in the main settings and reaches 3.18x speedup at 0.25\textdegree{}, with over 2.2x lower peak memory. A controlled decomposition shows that the accuracy gain comes primarily from sparse routing itself, while random token selection provides an additional regularization benefit without selector overhead.

Figures

Figures reproduced from arXiv: 2607.02829 by Janet Wang, Jihun Hamm, Lin Zhao, Xiao Wang, Xi Xiao, Yunbei Zhang.

Figure 1
Figure 1. Figure 1: Overview of Sparse-Reslim. (a) Three-stage token routing: dense early (n1 blocks), sparse middle (n2 blocks on K=⌊rL⌋ tokens), and dense late (n3 blocks), with unselected tokens following an identity path. (b, c) Integration into the deterministic Res-Slim-ViT and generative Res-Slim-EDM backbones. a multi-modal, multi-task cross-modal fusion transformer to push skillful fore￾casts beyond 10 lead days, and… view at source ↗
Figure 2
Figure 2. Figure 2: Training convergence for Res-Slim-ViT at 1.40625° and 1.0°. Sparse-Reslim (solid) converges faster than the dense baseline (dashed) in both wall-clock time and training loss, reaching lower final values at both resolutions. tocol of ClimaX [17], we use 1979–2017 for training, 2018–2019 for validation, and 2020 for testing. The input consists of 40 atmospheric variables: six vari￾ables (geopotential, temper… view at source ↗
Figure 3
Figure 3. Figure 3: Training efficiency across resolutions for the deterministic model. Sparse-Reslim achieves consistent wall-clock speedup at both resolutions, with larger gains at 1.0° due to the higher token count (L=16,200 vs. 8,192). Implementation details. All models are trained with Fully Sharded Data Parallel (FSDP) and activation checkpointing. The deterministic model runs on 2 nodes with 16 AMD MI250X GPUs, using A… view at source ↗
Figure 4
Figure 4. Figure 4: Forecast quality of the EDM generative model on the ERA5 test set (year 2020, 120-hour lead time) at 1.0°. Sparse-Reslim uses r=0.25 with block split (2, 4, 2). Bold = better. Variable Method RMSE↓ ACC↑ T2m Dense EDM 3.510 0.771 Sparse-Reslim EDM 3.280 0.802 U10 Dense EDM 4.420 0.204 Sparse-Reslim EDM 4.170 0.251 Z500 Dense EDM 951.300 0.517 Sparse-Reslim EDM 889.700 0.571 T850 Dense EDM 4.220 0.596 Sparse… view at source ↗
Figure 5
Figure 5. Figure 5: Keep ratio r vs. Z500 RMSE and speedup. r=0.25 is the sweet spot. Effect of block split (n1, n2, n3) [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Scaling behavior: both speedup and RMSE improvement grow with model size. We introduced Sparse-Reslim, a parameter￾free sparse token routing module that partitions ViT blocks into dense early, sparse middle, and dense late stages, processing only a random 25% of tokens in the dominant middle lay￾ers while a residual delta formulation maintains spatially dense outputs. Experiments on ERA5 at three res￾oluti… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 4 linked inside Pith

  1. [1]

    Nature619(7970), 533–538 (2023) 2, 3, 23

    Bi, K., Xie, L., Zhang, H., Chen, X., Gu, X., Tian, Q.: Accurate medium-range global weather forecasting with 3d neural networks. Nature619(7970), 533–538 (2023) 2, 3, 23

    2023

  2. [2]

    Nature641(8065), 1180–1187 (2025) 2, 4

    Bodnar, C., Bruinsma, W.P., Lucic, A., Stanley, M., Allen, A., Brandstetter, J., Garvan, P., Riechert, M., Weyn, J.A., Dong, H., et al.: A foundation model for the earth system. Nature641(8065), 1180–1187 (2025) 2, 4

    2025

  3. [3]

    arXiv preprint arXiv:2210.09461 (2022) 2, 4

    Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022) 2, 4

    Pith/arXiv arXiv 2022

  4. [4]

    arXiv preprint arXiv:2304.02948 (2023) 3

    Chen, K., Han, T., Gong, J., Bai, L., Ling, F., Luo, J.J., Chen, X., Ma, L., Zhang, T., Su, R., et al.: Fengwu: Pushing the skillful global medium-range weather fore- cast beyond 10 days lead. arXiv preprint arXiv:2304.02948 (2023) 3

    arXiv 2023

  5. [5]

    npj climate and atmospheric science6(1), 190 (2023) 2, 4

    Chen, L., Zhong, X., Zhang, F., Cheng, Y., Xu, Y., Qi, Y., Li, H.: Fuxi: a cascade machine learning forecasting system for 15-day global weather forecast. npj climate and atmospheric science6(1), 190 (2023) 2, 4

    2023

  6. [6]

    Advances in neural information processing systems34, 19974–19988 (2021) 5

    Chen, T., Cheng, Y., Gan, Z., Yuan, L., Zhang, L., Wang, Z.: Chasing sparsity in vision transformers: An end-to-end exploration. Advances in neural information processing systems34, 19974–19988 (2021) 5

    2021

  7. [7]

    Geoscientific Model Development11(10), 3999–4009 (2018) 1

    Dueben, P.D., Bauer, P.: Challenges and design choices for global weather and cli- mate models based on machine learning. Geoscientific Model Development11(10), 3999–4009 (2018) 1

    2018

  8. [8]

    In: European conference on computer vision

    Fayyaz, M., Koohpayegani, S.A., Jafari, F.R., Sengupta, S., Joze, H.R.V., Som- merlade, E., Pirsiavash, H., Gall, J.: Adaptive token sampling for efficient vision transformers. In: European conference on computer vision. pp. 396–414. Springer (2022) 4

    2022

  9. [9]

    Copernicus climate change service (c3s) climate data store (cds)10(1.24381), 24381 (2023) 9

    Hersbach, H., Bell, B., Berrisford, P., Biavati, G., Horányi, A., Muñoz Sabater, J., Nicolas, J., Peubey, C., Radu, R., Rozum, I., et al.: Era5 hourly data on single levels from 1940 to present. Copernicus climate change service (c3s) climate data store (cds)10(1.24381), 24381 (2023) 9

    1940

  10. [10]

    Quarterly journal of the royal meteorological society146(730), 1999–2049 (2020) 3, 9, 23

    Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Horányi, A., Muñoz-Sabater, J., Nicolas, J., Peubey, C., Radu, R., Schepers, D., et al.: The era5 global reanalysis. Quarterly journal of the royal meteorological society146(730), 1999–2049 (2020) 3, 9, 23

    1999

  11. [11]

    Advances in neural information processing systems35, 26565–26577 (2022) 3, 6, 10, 20

    Karras,T.,Aittala,M.,Aila,T.,Laine,S.:Elucidatingthedesignspaceofdiffusion- based generative models. Advances in neural information processing systems35, 26565–26577 (2022) 3, 6, 10, 20

    2022

  12. [12]

    In: European conference on computer vision

    Kong, Z., Dong, P., Ma, X., Meng, X., Niu, W., Sun, M., Shen, X., Yuan, G., Ren, B.,Tang,H.,etal.:Spvit:Enablingfastervisiontransformersvialatency-awaresoft token pruning. In: European conference on computer vision. pp. 620–640. Springer (2022) 2, 4

    2022

  13. [13]

    Science382(6677), 1416–1421 (2023) 2, 3, 23

    Lam, R., Sanchez-Gonzalez, A., Willson, M., Wirnsberger, P., Fortunato, M., Alet, F., Ravuri, S., Ewalds, T., Eaton-Rosen, Z., Hu, W., et al.: Learning skillful medium-range global weather forecasting. Science382(6677), 1416–1421 (2023) 2, 3, 23

    2023

  14. [14]

    arXiv preprint arXiv:2202.07800 (2022) 2, 4

    Liang, Y., Ge, C., Tong, Z., Song, Y., Wang, J., Xie, P.: Not all patches are what you need: Expediting vision transformers via token reorganizations. arXiv preprint arXiv:2202.07800 (2022) 2, 4

    Pith/arXiv arXiv 2022

  15. [15]

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer:Hierarchicalvisiontransformerusingshiftedwindows.In:Proceedings Sparse Residual Routing for Weather Prediction 17 of the IEEE/CVF international conference on computer vision. pp. 10012–10022 (2021) 2, 5

    2021

  16. [16]

    Communications Earth & Envi- ronment6(1), 124 (2025) 2, 4

    Mardani, M., Brenowitz, N., Cohen, Y., Pathak, J., Chen, C.Y., Liu, C.C., Vahdat, A., Nabian, M.A., Ge, T., Subramaniam, A., et al.: Residual corrective diffusion modeling for km-scale atmospheric downscaling. Communications Earth & Envi- ronment6(1), 124 (2025) 2, 4

    2025

  17. [17]

    arXiv preprint arXiv:2301.10343 (2023) 2, 3, 10

    Nguyen, T., Brandstetter, J., Kapoor, A., Gupta, J.K., Grover, A.: Climax: A foundation model for weather and climate. arXiv preprint arXiv:2301.10343 (2023) 2, 3, 10

    Pith/arXiv arXiv 2023

  18. [18]

    arXiv preprint arXiv:2510.18707 (2025) 4

    Nguyen, T., Pham, T., Arcomano, T., Kotamarthi, V., Foster, I., Madireddy, S., Grover, A.: Omnicast: A masked latent diffusion model for weather forecasting across time scales. arXiv preprint arXiv:2510.18707 (2025) 4

    arXiv 2025

  19. [19]

    Advances in Neural Information Processing Systems37, 68740–68771 (2024) 2, 3

    Nguyen, T., Shah, R., Bansal, H., Arcomano, T., Maulik, R., Kotamarthi, R., Fos- ter, I., Madireddy, S., Grover, A.: Scaling transformer neural networks for skillful and reliable medium-range weather forecasting. Advances in Neural Information Processing Systems37, 68740–68771 (2024) 2, 3

    2024

  20. [20]

    arXiv preprint arXiv:2510.21986 (2025) 2, 5, 7

    Park, D., Haji-Ali, M., Li, Y., Menapace, W., Tulyakov, S., Kim, H.J., Siarohin, A., Kag, A.: Sprint: Sparse-dense residual fusion for efficient diffusion transformers. arXiv preprint arXiv:2510.21986 (2025) 2, 5, 7

    arXiv 2025

  21. [21]

    arXiv preprint arXiv:2312.15796 (2023) 2, 4

    Price,I.,Sanchez-Gonzalez,A.,Alet,F.,Andersson,T.R.,El-Kadi,A.,Masters,D., Ewalds, T., Stott, J., Mohamed, S., Battaglia, P., et al.: Gencast: Diffusion-based ensemble forecasting for medium-range weather. arXiv preprint arXiv:2312.15796 (2023) 2, 4

    Pith/arXiv arXiv 2023

  22. [22]

    IEEE Transactions on Pattern Analysis and Machine Intelligence45(9), 10883–10897 (2023) 4

    Rao, Y., Liu, Z., Zhao, W., Zhou, J., Lu, J.: Dynamic spatial sparsification for efficient vision transformers and convolutional neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence45(9), 10883–10897 (2023) 4

    2023

  23. [23]

    Advances in neural infor- mation processing systems34, 13937–13949 (2021) 2, 4

    Rao, Y., Zhao, W., Liu, B., Lu, J., Zhou, J., Hsieh, C.J.: Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural infor- mation processing systems34, 13937–13949 (2021) 2, 4

    2021

  24. [24]

    Journal of Ad- vances in Modeling Earth Systems12(11), e2020MS002203 (2020) 9

    Rasp, S., Dueben, P.D., Scher, S., Weyn, J.A., Mouatadid, S., Thuerey, N.: Weath- erbench: a benchmark data set for data-driven weather forecasting. Journal of Ad- vances in Modeling Earth Systems12(11), e2020MS002203 (2020) 9

    2020

  25. [25]

    Journal of Advances in Modeling Earth Systems16(6), e2023MS004019 (2024) 9

    Rasp, S., Hoyer, S., Merose, A., Langmore, I., Battaglia, P., Russell, T., Sanchez- Gonzalez, A., Yang, V., Carver, R., Agrawal, S., et al.: Weatherbench 2: A bench- mark for the next generation of data-driven global weather models. Journal of Advances in Modeling Earth Systems16(6), e2023MS004019 (2024) 9

    2024

  26. [26]

    Geophysical Research Letters 45(22), 12–616 (2018) 1

    Scher, S.: Toward data-driven weather and climate forecasting: Approximating a simple general circulation model with deep learning. Geophysical Research Letters 45(22), 12–616 (2018) 1

    2018

  27. [27]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2025) 24

    Wang, J., Zhang, Y., Ding, Z., Hamm, J.: Doctor approved: Generating medically accurate skin disease images through ai-expert feedback. In: Advances in Neural Information Processing Systems (NeurIPS) (2025) 24

    2025

  28. [28]

    In: Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis

    Wang, X., Choi, J.Y., Kurihaya, T., Lyngaas, I., Yoon, H.J., Xiao, X., Pugmire, D., Fan, M., Nafi, N.M., Tsaris, A., et al.: Orbit-2: Scaling exascale vision foundation models for weather and climate downscaling. In: Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis. pp. 86–98 (2025) 3, 4, 10

    2025

  29. [29]

    In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition

    Wei, C., Duke, B., Jiang, R., Aarabi, P., Taylor, G.W., Shkurti, F.: Sparsifiner: Learning sparse instance-dependent attention for efficient vision transformers. In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition. pp. 22680–22689 (2023) 5 18 J. Wang et al

    2023

  30. [30]

    Journal of Advances in Modeling Earth Systems11(8), 2680–2693 (2019) 1

    Weyn, J.A., Durran, D.R., Caruana, R.: Can machines learn to predict weather? using deep learning to predict gridded 500-hpa geopotential height from historical weather data. Journal of Advances in Modeling Earth Systems11(8), 2680–2693 (2019) 1

    2019

  31. [31]

    In: Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Xiao, X., Ma, C., Zhang, Y., Liu, C., Wang, Z., Li, Y., Zhao, L., Hu, G., Wang, T., Xu, H.: Not all directions matter: Towards structured and task-aware low-rank model adaptation. In: Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2132–2154 (2026) 24

    2026

  32. [32]

    arXiv preprint arXiv:2510.13219 (2025) 24

    Xiao, X., Zhang, Y., Zhao, L., Liu, Y., Liao, X., Mai, Z., Li, X., Wang, X., Xu, H., Hamm, J., et al.: Prompt-based adaptation in large-scale vision models: A survey. arXiv preprint arXiv:2510.13219 (2025) 24

    arXiv 2025

  33. [33]

    In: Proceedings of the AAAI conference on artificial intelligence

    Xu, Y., Zhang, Z., Zhang, M., Sheng, K., Li, K., Dong, W., Zhang, L., Xu, C., Sun, X.: Evo-vit: Slow-fast token evolution for dynamic vision transformer. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2964– 2972 (2022) 2, 4

    2022

  34. [34]

    OpenReview (2025), https://openreview.net/forum?id=Hic8PwzlBY24

    Zhang, Y., Cai, C., Liu, F., Hamm, J.: Prime Once, then Reprogram Locally: An Efficient Alternative to Black-Box Model Reprogramming. OpenReview (2025), https://openreview.net/forum?id=Hic8PwzlBY24

    2025

  35. [35]

    In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

    Zhang, Y., Mehra, A., Hamm, J.: Ot-vp: Optimal transport-guided visual prompt- ing for test-time adaptation. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1122–1132. IEEE (2025), oral Presentation 24

    2025

  36. [36]

    In: Forty-second International Conference on Machine Learning (ICML) (2025) 24

    Zhang, Y., Mehra, A., Niu, S., Hamm, J.: Dpcore: Dynamic prompt coreset for con- tinual test-time adaptation. In: Forty-second International Conference on Machine Learning (ICML) (2025) 24

    2025

  37. [37]

    number of training steps

    Zhang, Y., Niu, S., Liu, F., Hamm, J.: Adapting in the Dark: Towards Sta- ble and Efficient Black-Box Test-Time Adaptation. OpenReview (2026),https: //openreview.net/forum?id=KBtySUdWuH24 Sparse Residual Routing for Weather Prediction 19 Appendix A Implementation Details A.1 Architecture Specifications Tables 6 and 7 consolidate the full architectural spe...

    2026