{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.27328", "kind": "paper", "url": "https://arxiv.org/abs/2608.27328"}, {"id": "gh:amap-ml/r2mbench", "kind": "repo", "url": "https://github.com/AMAP-ML/R2MBench"}], "authors": ["Qiwen Gu", "Bingjie Gao", "Rui Chen", "Geng Li", "Jifan Li", "Qishuai Wen", "Li Niu", "Jing Tang", "Xiangxiang Chu", "Junqiao Zhao"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.27328", "source_url": "https://arxiv.org/abs/2608.27328"}, "categories": ["General AI", "Vision & 3D", "cs.CV"], "description": {"en": "Evaluates revisit memory in interactive video world models using relative consistency: for each detected return, compares revisit pair with gap-matched non-revisit and short-range control pairs to compute MemoryGain and Normalized Memory Ratio across appearance, identity, geometry, and state."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.27328", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.27328", "title": "R2M-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Tongji University", "Alibaba Group (China)", "Shanghai Jiao Tong University"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2608.27328", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2608.27328", "source_url": "https://openalex.org/W7204566455"}, "key": "claire-radar:2608.27328", "modality": null, "name": "R2M-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": "true", "origin_record_id": "bm_r2m-bench_1ea8e143", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "5df8e50e9487d95d56a77c71ca13714b16a43f7157ca69db610d9602b11b6d5f", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-08T08:58:25.506689Z", "source_url": "https://arxiv.org/abs/2608.27328"}, "publisher": null, "released": "2026-08-27", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.27328", "source_url": "https://arxiv.org/abs/2608.27328"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-27328", "source": "claire_radar", "source_benchmark_id": "2608.27328", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Vision & 3D", "attention": {"asOf": "2026-09-08", "attemptedAt": "2026-09-08T01:38:09.842216Z", "githubRepo": "https://github.com/AMAP-ML/R2MBench", "githubScope": "benchmark_repo", "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2608.27328", "observedAt": "2026-09-07T01:28:43.988909Z", "signalStatus": {"githubStars": {"checkedAt": "2026-09-08T01:38:09.842216Z", "state": "unavailable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-09-08T01:38:09.842216Z", "state": "unavailable"}}, "status": "unavailable"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The benchmark addresses a specific evaluation gap in video world models, offers clear metrics and code, and could attract interest from researchers in video generation and world models, though the topic is relatively niche.", "score": 65}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-08T08:58:25.506689Z", "inputHash": "ac4c0a87d2242184cfcafd417b6d672dc690b9017921a3a603422b8bcd747b4c", "model": "deepseek-v4-pro"}, "curation": {"canonicalNameEvidence": "We introduce \\emph{R2M-Bench} (\\textbf{R}elative \\textbf{R}evisit \\textbf{M}emory Benchmark)", "canonicalNameSource": "abstract", "decisionReason": "Benchmark has a formal name, defined scoring metrics, 300 instances, evaluates multiple models, and provides code link for reproducibility.", "model": "deepseek-v4-pro", "reviewedAt": "2026-09-08T08:58:25.506689Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates revisit memory in interactive video world models using relative consistency: for each detected return, compares revisit pair with gap-matched non-revisit and short-range control pairs to compute MemoryGain and Normalized Memory Ratio across appearance, identity, geometry, and state.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "We introduce \\emph{R2M-Bench} (\\textbf{R}elative \\textbf{R}evisit \\textbf{M}emory Benchmark), a benchmark of observable revisit-selective consistency."}, "familyId": "bmf_4bb72cc2e03a", "firstSeenAt": "2026-08-30", "heat": null, "id": "bm_r2m-bench_1ea8e143", "indexedAt": "2026-08-30T06:59:55.242506Z", "industrySectors": [], "links": {"code": "https://github.com/AMAP-ML/R2MBench", "data": null, "hfPaper": "https://huggingface.co/papers/2608.27328", "paper": "https://arxiv.org/abs/2608.27328", "pdf": "https://arxiv.org/pdf/2608.27328", "project": null, "report": "https://arxiv.org/abs/2608.27328"}, "metrics": [], "motivation": "High similarity between first-visit and return frames does not necessarily show that a video world model remembered the scene; the intervening rollout may simply have changed very little.", "name": "R2M-Bench", "oneLine": "Evaluates revisit memory in interactive video world models using relative consistency: for each detected return, compares revisit pair with gap-matched non-revisit and short-range control pairs to compute MemoryGain and Normalized Memory Ratio across appearance, identity, geometry, and state.", "paperTitle": "R2M-Bench: Evaluating Revisit Memory via Relative Consistency in Interactive Video World Models", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.27328", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-08T06:12:53.942187Z"}, "publications": [], "publishers": [{"name": "AMAP-ML", "organizationType": "company-research-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/AMAP-ML/R2MBench"}], "ranking": {"30d": {"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": 30}, "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}, "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": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-08-27", "source": {"authors": ["Qiwen Gu", "Bingjie Gao", "Rui Chen", "Geng Li", "Jifan Li", "Qishuai Wen", "Li Niu", "Jing Tang", "Xiangxiang Chu", "Junqiao Zhao"], "categories": ["cs.CV"], "id": "2608.27328", "title": "R2M-Bench: Evaluating Revisit Memory via Relative Consistency in Interactive Video World Models", "type": "arxiv", "url": "https://arxiv.org/abs/2608.27328"}, "sourceUpdatedAt": "2026-08-27T16:26:29Z", "topics": ["cs.CV"], "venueAttempts": [], "whyItMatters": "Provides a calibration method that isolates revisit-specific consistency from generic temporal stability, addressing slow-motion shortcut and enabling more meaningful comparison of video world models' memory capabilities."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
