{"paper":{"title":"(Sparse) Attention to the Details: Preserving Spectral Fidelity in ML-based Weather Forecasting Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Mosaic achieves near-perfect spectral alignment in 1.5° weather forecasts by using block-sparse attention and learned ensemble perturbations, matching finer-resolution models.","cross_cats":["cs.AI","cs.CV","physics.ao-ph"],"primary_cat":"cs.LG","authors_text":"Ana Lucic, Jan-Willem van de Meent, Maksim Zhdanov, Max Welling","submitted_at":"2026-04-06T08:50:42Z","abstract_excerpt":"We introduce Mosaic, a probabilistic weather forecasting model that addresses three failure modes of spectral degradation in ML-based weather prediction: spectral damping (statistical), high-frequency aliasing (architectural), and residual high-frequency leakage (parametric). Mosaic generates ensemble members through learned functional perturbations and operates on native-resolution grids via mesh-aligned block-sparse attention, a hardware-aligned mechanism that captures long-range dependencies at linear cost by sharing keys and values across spatially adjacent queries. At 1.5{\\deg} resolution"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Mosaic produces well-calibrated ensembles whose individual members exhibit near-perfect spectral alignment across all resolved frequencies at 1.5° resolution while matching or outperforming models trained on 6× finer grids.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the mesh-aligned block-sparse attention fully captures necessary long-range dependencies without introducing new artifacts or losing critical interactions that standard attention would preserve.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Mosaic achieves state-of-the-art spectral alignment in 1.5° weather forecasts via learned functional perturbations and hardware-aligned sparse attention, matching finer-resolution models with fast inference.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Mosaic achieves near-perfect spectral alignment in 1.5° weather forecasts by using block-sparse attention and learned ensemble perturbations, matching finer-resolution models.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"d1ea6d4dafd37b86feaba14960ed678d8150e8bed21341459835f0c385df26a5"},"source":{"id":"2604.16429","kind":"arxiv","version":3},"verdict":{"id":"cd44595f-8101-42c0-980a-4cbec0767b67","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T21:13:35.022788Z","strongest_claim":"Mosaic produces well-calibrated ensembles whose individual members exhibit near-perfect spectral alignment across all resolved frequencies at 1.5° resolution while matching or outperforming models trained on 6× finer grids.","one_line_summary":"Mosaic achieves state-of-the-art spectral alignment in 1.5° weather forecasts via learned functional perturbations and hardware-aligned sparse attention, matching finer-resolution models with fast inference.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the mesh-aligned block-sparse attention fully captures necessary long-range dependencies without introducing new artifacts or losing critical interactions that standard attention would preserve.","pith_extraction_headline":"Mosaic achieves near-perfect spectral alignment in 1.5° weather forecasts by using block-sparse attention and learned ensemble perturbations, matching finer-resolution models."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.16429/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"0d4873aeb516f38615bdd0522eebdc9cd70bfd65afef0f54aaf12ac19664c4d5"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}