{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.21839", "kind": "paper", "url": "https://arxiv.org/abs/2608.21839v1"}], "authors": ["Peiyuan Zhang", "Xiangyu Zhao", "Hongbo Liu", "Xiaoxing Hu", "Mingxin Liu", "Shuran Ma", "Yunhang Shen", "Jian Hu", "Haihan Gao", "Haoyu Cao", "Xue Yang"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.21839", "source_url": "https://arxiv.org/abs/2608.21839"}, "categories": ["General AI", "Vision & 3D", "cs.CV"], "description": {"en": "FIRM-Video-Bench is a benchmark of 250 text-to-video generations with 750 point-wise human annotations covering Instruction Following, World Coherence, and Perceptual Quality."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.21839v1", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.21839v1", "title": "FIRM-Video-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2608.21839", "modality": null, "name": "FIRM-Video-Bench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_firm-video-bench_74b2abde", "origin_source": "arxiv", "recognition_confidence": "0.65", "record_sha256": "fef93c4ae919c2c8791a7590fc3cf9f1516b530817b3276e96d9069623a4bb1b", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-08-25T17:38:30.498900Z", "source_url": "https://arxiv.org/abs/2608.21839v1"}, "publisher": null, "released": "2026-08-22", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.21839", "source_url": "https://arxiv.org/abs/2608.21839v1"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-21839", "source": "claire_radar", "source_benchmark_id": "2608.21839", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Mixed", "applicationDomains": ["General AI"], "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": 4, "hfPaperUrl": "https://huggingface.co/papers/2608.21839", "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": "2026-09-30", "state": "stale"}}, "status": "stale"}, "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "A narrowly scoped text-to-video reward modeling benchmark with a moderate dataset size and no independent adoption signals is likely to draw limited immediate attention.", "score": 15}, "capabilities": [], "confidence": "Medium", "construction": "Aggregate Existing", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-25T17:38:30.498900Z", "inputHash": "c3567366b3b4b292c0b491576fc1c7b91b973339abdee8aa94ac64a2299036e6", "model": "deepseek-v4-pro"}, "curation": {"canonicalNameEvidence": "introduce FIRM-Video-Bench with 750 point-wise human annotations across 250 videos", "canonicalNameSource": "abstract", "decisionReason": "The abstract formally names FIRM-Video-Bench and specifies a public dataset with human annotations, satisfying a reusable evaluation object and scoring contract.", "model": "deepseek-v4-pro", "reviewedAt": "2026-08-25T17:38:30.498900Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "FIRM-Video-Bench is a benchmark of 250 text-to-video generations with 750 point-wise human annotations covering Instruction Following, World Coherence, and Perceptual Quality.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "evaluation protocol evidence"], "snippet": "Subsequently, we construct FIRM-Video-90K with 88,044 dimension-specific instances from 29,348 videos, and introduce FIRM-Video-Bench with 750 point-wise human annotations across 250 videos."}, "familyId": "bmf_42680619c993", "firstSeenAt": "2026-08-25", "heat": null, "id": "bm_firm-video-bench_74b2abde", "indexedAt": "2026-08-25T17:37:25.889905Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "http://arxiv.org/abs/2608.21839v1", "pdf": "https://arxiv.org/pdf/2608.21839v1", "project": null, "report": "http://arxiv.org/abs/2608.21839v1"}, "metrics": [], "motivation": "Reliable reward models are essential for text-to-video evaluation and alignment.", "name": "FIRM-Video-Bench", "oneLine": "FIRM-Video-Bench is a benchmark of 250 text-to-video generations with 750 point-wise human annotations covering Instruction Following, World Coherence, and Perceptual Quality.", "paperTitle": "FIRM-Video: Check Before You Score for Reliable Text-to-Video Reward Modeling", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "http://arxiv.org/abs/2608.21839v1", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-25T17:37:25.889905Z"}, "publications": [], "ranking": {"90d": {"confidence": "Low", "coverage": 0.15, "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": 0.2843790034276272, "value": 4}}, "confidence": "Low", "coverage": 0.15, "rank": null, "score": 47}, "method": "attention-ranking-v13", "rank": null, "score": 47}}, "readiness": "Paper only", "recognitionConfidence": 0.65, "relation": "introduces", "releasedAt": "2026-08-22", "source": {"authors": ["Peiyuan Zhang", "Xiangyu Zhao", "Hongbo Liu", "Xiaoxing Hu", "Mingxin Liu", "Shuran Ma", "Yunhang Shen", "Jian Hu", "Haihan Gao", "Haoyu Cao", "Xue Yang"], "categories": ["cs.CV"], "id": "2608.21839", "title": "FIRM-Video: Check Before You Score for Reliable Text-to-Video Reward Modeling", "type": "arxiv", "url": "http://arxiv.org/abs/2608.21839v1"}, "sourceUpdatedAt": "2026-08-22T08:21:40Z", "topics": ["cs.CV"], "venueAttempts": [], "whyItMatters": "It provides a fixed annotated set for evaluating reward models' ability to score videos on fine-grained dimensions, enabling comparison of text-to-video evaluation approaches."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
