{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.13167", "kind": "paper", "url": "https://arxiv.org/abs/2608.13167"}], "authors": ["Fnu Pramono", "John Cai", "Sourabh Kulkarni"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.13167", "source_url": "https://arxiv.org/abs/2608.13167"}, "categories": ["General AI", "Multimodal"], "description": {"en": "TRAPSBench is a procedurally generated video benchmark with 1,404 physics pairs to test epistemic restraint in VLMs, using Penalized Epistemic Calibration Score (PECS)."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.13167", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.13167", "title": "TRAPSBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Meta (United Kingdom)"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2608.13167", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2608.13167", "source_url": "https://openalex.org/W7203470611"}, "key": "claire-radar:2608.13167", "modality": null, "name": "TRAPSBench", "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_trapsbench_2047e708", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "c8e47c9b5bdcbf2c9f2872157fbcef7e8da04eda1fd34908ce404bc797b595a3", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2608.13167"}, "publisher": null, "released": "2026-08-13", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.13167", "source_url": "https://arxiv.org/abs/2608.13167"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-13167", "source": "claire_radar", "source_benchmark_id": "2608.13167", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Machine Generated", "applicationDomains": ["General AI"], "area": "Multimodal", "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/2608.13167", "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": "Original Synthetic", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "308917b653506a73ef52456ad12533ea709e909f580053828fc58540113ff1d5", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "TRAPSBench is a procedurally generated video benchmark with 1,404 physics pairs to test epistemic restraint in VLMs, using Penalized Epistemic Calibration Score (PECS).", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We introduce TRAPSBench, a procedurally generated video benchmark of 1,404 matched physics pairs in which a single targeted change renders the outcome undeterminable from the visual evidence."}, "familyId": "bmf_bffec585a3f1", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_trapsbench_2047e708", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2608.13167", "paper": "https://arxiv.org/abs/2608.13167", "pdf": "https://arxiv.org/pdf/2608.13167", "project": null, "report": "https://arxiv.org/abs/2608.13167"}, "metrics": [], "motivation": "When visual evidence is occluded or chaotic, models should abstain.", "name": "TRAPSBench", "oneLine": "TRAPSBench is a procedurally generated video benchmark with 1,404 physics pairs to test epistemic restraint in VLMs, using Penalized Epistemic Calibration Score (PECS).", "paperTitle": "TRAPSBench: Vision-Language Models Encode but Fail to Express Epistemic Restraint", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.13167", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {"90d": {"confidence": "Low", "coverage": 0.0, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 90}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": null, "value": null}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v13", "rank": null, "score": null}}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-08-13", "source": {"authors": ["Fnu Pramono", "John Cai", "Sourabh Kulkarni"], "categories": ["cs.CV", "cs.AI", "cs.CL", "cs.LG"], "id": "2608.13167", "title": "TRAPSBench: Vision-Language Models Encode but Fail to Express Epistemic Restraint", "type": "arxiv", "url": "https://arxiv.org/abs/2608.13167"}, "sourceUpdatedAt": "2026-08-13T12:32:13Z", "topics": ["Multimodal"], "venueAttempts": [], "whyItMatters": "It highlights that VLMs can internally detect when abstention is required but fail to express it, offering a metric for calibration and restraint evaluation."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
