{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.30201", "kind": "paper", "url": "https://arxiv.org/abs/2606.30201"}], "authors": ["Filippo Ruffini", "Marco Salmé", "Rosa Sicilia", "Valerio Guarrasi", "Paolo Soda"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.30201", "source_url": "https://arxiv.org/abs/2606.30201"}, "categories": ["Biology & Drug Discovery", "Vision & 3D", "cs.CV"], "description": {"en": "SHOVIR evaluates vision shortcut learning in radiology report generation by extending MIMIC-CXR and PadChest-GR with per-box CheXpert labels. It defines image-level and disease-level occlusion experiments that compare model predictions on clean images against localized perturbations to isolate direct and contextual shortcut failures at the disease-class level."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.30201", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.30201", "title": "SHOVIR"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Università Campus Bio-Medico", "Umeå University", "Saint Camillus International University of Health and Medical Sciences"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.30201", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.30201", "source_url": "https://openalex.org/W7166740925"}, "key": "claire-radar:2606.30201", "modality": null, "name": "SHOVIR", "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_shovir_1cca6b61", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "dc0e2188641c5a861838f3857536c12b589ba76bd7023755bfd668c0fc489057", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.30201"}, "publisher": null, "released": "2026-06-29", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.30201", "source_url": "https://arxiv.org/abs/2606.30201"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-30201", "source": "claire_radar", "source_benchmark_id": "2606.30201", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Mixed", "applicationDomains": ["Biology & Drug Discovery"], "area": "Vision & 3D", "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": 0, "hfPaperUrl": "https://huggingface.co/papers/2606.30201", "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", "lastSuccessfulDate": "2026-10-04", "state": "stale"}}, "status": "stale"}, "capabilities": [], "confidence": "High", "construction": "Aggregate Existing", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "222d6f65b9d787c3b317be7a0386db381e68803608001c41ba22a0832d825ce7", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "SHOVIR evaluates vision shortcut learning in radiology report generation by extending MIMIC-CXR and PadChest-GR with per-box CheXpert labels. It defines image-level and disease-level occlusion experiments that compare model predictions on clean images against localized perturbations to isolate direct and contextual shortcut failures at the disease-class level.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We introduce SHOVIR, a benchmark for evaluating vision shortcut behavior in RRG."}, "familyId": "bmf_72ae455766ea", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_shovir_1cca6b61", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Pharma & Biotech"], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2606.30201", "paper": "https://arxiv.org/abs/2606.30201", "pdf": "https://arxiv.org/pdf/2606.30201", "project": null, "report": "https://arxiv.org/abs/2606.30201"}, "metrics": [], "motivation": "Current evaluation protocols for Vision-Language Models (VLMs) in Radiology Report Generation (RRG) rely on report-level metrics that measure lexical overlap or aggregate clinical correctness.", "name": "SHOVIR", "oneLine": "SHOVIR evaluates vision shortcut learning in radiology report generation by extending MIMIC-CXR and PadChest-GR with per-box CheXpert labels. It defines image-level and disease-level occlusion experiments that compare model predictions on clean images against localized perturbations to isolate direct and contextual shortcut failures at the disease-class level.", "paperTitle": "SHOVIR: A Benchmark for Evaluating Vision Shortcut Learning in Radiology Report Generation", "primaryDomain": "Biology & Drug Discovery", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.30201", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-06-29", "source": {"authors": ["Filippo Ruffini", "Marco Salm\\'e", "Rosa Sicilia", "Valerio Guarrasi", "Paolo Soda"], "categories": ["cs.CV", "cs.CL"], "id": "2606.30201", "title": "SHOVIR: A Benchmark for Evaluating Vision Shortcut Learning in Radiology Report Generation", "type": "arxiv", "url": "https://arxiv.org/abs/2606.30201"}, "sourceUpdatedAt": "2026-07-22T13:09:26Z", "topics": ["cs.CV"], "venueAttempts": [], "whyItMatters": "Standard RRG metrics miss whether diagnostic statements are grounded in actual image evidence, allowing models to exploit dataset priors. SHOVIR provides a protocol to assess spatial grounding, revealing that high report quality can coexist with shallow visual reliance, which is critical for clinical deployment decisions."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
