{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2605.24782", "kind": "paper", "url": "https://arxiv.org/abs/2605.24782"}], "authors": ["Dingling Yao", "Andrea Polesello", "Adeel Pervez", "Caroline Muller", "Francesco Locatello"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2605.24782", "source_url": "https://arxiv.org/abs/2605.24782"}, "categories": ["General AI", "Language & Knowledge", "Safety"], "description": {"en": "A benchmark dataset for tropical cyclone research with an automated construction pipeline, used to probe scientific alignment of vision foundation models."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2605.24782", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2605.24782", "title": "TC-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2605.24782", "modality": null, "name": "TC-Bench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "High", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "unknown", "origin_record_id": "bm_tc-bench_6009a2aa", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "5f4a95dcb60532288f2871542b53f6374a670957b7a0d4e57a4de7ba813d34c8", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2605.24782"}, "publisher": null, "released": "2026-05-23", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2605.24782", "source_url": "https://arxiv.org/abs/2605.24782"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2605-24782", "source": "claire_radar", "source_benchmark_id": "2605.24782", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2605.24782", "observedAt": "2026-10-04T03:02:42.754940Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "state": "unavailable"}}, "status": "unavailable"}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "488472c69ff7d173ad64263cf6b3f92eae30f62ad5092da7cfa42c4932096a96", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A benchmark dataset for tropical cyclone research with an automated construction pipeline, used to probe scientific alignment of vision foundation models.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "To operationalize this framework, we release TC-Bench, a global, reproducible benchmark dataset with an automated construction pipeline for tropical cyclone research, and show that current VFMs rely on visual shortcuts that collapse in intense regimes, indicating that scientific alignment does not arise as a natural byproduct of scaling alone."}, "familyId": "bmf_4b558e2721e0", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_tc-bench_6009a2aa", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2605.24782", "paper": "https://arxiv.org/abs/2605.24782", "pdf": "https://arxiv.org/pdf/2605.24782", "project": null, "report": "https://arxiv.org/abs/2605.24782"}, "metrics": [], "motivation": "While Vision Foundation Models (VFMs) excel at predictive tasks on satellite imagery, their performance can arise from visual correlations rather than underlying structural invariants, making even perception-based out-of-distribution accuracy a poor proxy for scientific utility.", "name": "TC-Bench", "oneLine": "A benchmark dataset for tropical cyclone research with an automated construction pipeline, used to probe scientific alignment of vision foundation models.", "paperTitle": "The Perception-Physics Paradox: Probing Scientific Alignment with TC-Bench", "primaryDomain": "General AI", "publication": {"evidence": "Accepted at ICML 2026", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2605.24782", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "ICML 2026", "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-05-23", "source": {"authors": ["Dingling Yao", "Andrea Polesello", "Adeel Pervez", "Caroline Muller", "Francesco Locatello"], "categories": ["cs.LG"], "id": "2605.24782", "title": "The Perception-Physics Paradox: Probing Scientific Alignment with TC-Bench", "type": "arxiv", "url": "https://arxiv.org/abs/2605.24782"}, "sourceUpdatedAt": "2026-06-02T19:00:16Z", "topics": ["Safety"], "venueAttempts": [{"decisionRaw": "Accepted at ICML 2026", "evidence": [{"level": "author-claim", "observedAt": "2026-10-14T01:46:58.058199Z", "rawValue": "Accepted at ICML 2026", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2605.24782"}], "reviewStatus": "accepted", "venueName": "ICML 2026"}], "whyItMatters": "Examines whether models rely on visual shortcuts rather than structural invariants in scientific domains."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
