{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.05747", "kind": "paper", "url": "https://arxiv.org/abs/2608.05747"}], "authors": ["Qifeng Zhang", "Kaixiang Huang", "Heng Dong", "Huang Fang", "Junting Chen", "Junjie Zhu", "Yonghang Chen", "Zhiyu Zhang", "Wei Li"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.05747", "source_url": "https://arxiv.org/abs/2608.05747"}, "categories": ["Robotics & Embodied AI", "Vision & 3D", "cs.CV"], "description": {"en": "GST-Bench is a VQA benchmark for global spatial intelligence in video understanding, covering synthetic videos and human-verified questions. It evaluates VLMs' ability to infer spatial relations from novel viewpoints and map egocentric observations to top-down views."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.05747", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.05747", "title": "GST-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2608.05747", "modality": null, "name": "GST-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_gst-bench_e194aeb7", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "35880017534ef54280ebbf1de2a7dfac826c1023129859936ed9162b92e1fd53", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2608.05747"}, "publisher": null, "released": "2026-08-06", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.05747", "source_url": "https://arxiv.org/abs/2608.05747"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-05747", "source": "claire_radar", "source_benchmark_id": "2608.05747", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Machine Generated", "applicationDomains": ["Robotics & Embodied AI"], "area": "Vision & 3D", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": "2026-08-07T00:00:00.000Z", "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 46, "hfPaperUrl": "https://huggingface.co/papers/2608.05747", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}}, "status": "fresh"}, "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": "67d2fc1f118b7574e621f0d8db9a585f23e343d8f2dc5b26b08bd9c23b8272a4", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "GST-Bench is a VQA benchmark for global spatial intelligence in video understanding, covering synthetic videos and human-verified questions. It evaluates VLMs' ability to infer spatial relations from novel viewpoints and map egocentric observations to top-down views.", "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": "To probe the cause of this gap, we construct GST-Bench-Local and find that models, despite strong local spatial understanding under the same task formulation, still fail to consolidate long-horizon observations into a globally consistent scene representation."}, "familyId": "bmf_dc90c37b15cc", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_gst-bench_e194aeb7", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Robotics"], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2608.05747", "paper": "https://arxiv.org/abs/2608.05747", "pdf": "https://arxiv.org/pdf/2608.05747", "project": null, "report": "https://arxiv.org/abs/2608.05747"}, "metrics": [], "motivation": "Spatial intelligence is fundamental to embodied agents, yet existing benchmarks focus on local spatial perception from single or few viewpoints, overlooking global spatial awareness over continuous, long-horizon visual streams.", "name": "GST-Bench", "oneLine": "GST-Bench is a VQA benchmark for global spatial intelligence in video understanding, covering synthetic videos and human-verified questions. It evaluates VLMs' ability to infer spatial relations from novel viewpoints and map egocentric observations to top-down views.", "paperTitle": "GST-Bench: Can VLMs Develop Global Spatial Awareness from Video?", "primaryDomain": "Robotics & Embodied AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.05747", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "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": 0.6803003269431226, "value": 46}}, "confidence": "Low", "coverage": 0.15, "rank": null, "score": 53}, "method": "attention-ranking-v13", "rank": null, "score": 53}}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-08-06", "source": {"authors": ["Qifeng Zhang", "Kaixiang Huang", "Heng Dong", "Huang Fang", "Junting Chen", "Junjie Zhu", "Yonghang Chen", "Zhiyu Zhang", "Wei Li"], "categories": ["cs.CV"], "id": "2608.05747", "title": "GST-Bench: Can VLMs Develop Global Spatial Awareness from Video?", "type": "arxiv", "url": "https://arxiv.org/abs/2608.05747"}, "sourceUpdatedAt": "2026-08-06T08:33:13Z", "topics": ["cs.CV"], "venueAttempts": [], "whyItMatters": "Existing video benchmarks focus on local spatial perception, while GST-Bench targets global spatial awareness over long-horizon videos, addressing a gap in evaluating embodied agents' spatial intelligence. It provides a scoring contract to compare VLMs and highlights a significant performance gap between models and humans."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
