{"record": {"aliases": ["Where to Look: Can Foundation Models Reach a Target Viewpoint Through Active Exploration?"], "artifacts": [{"id": "arxiv:2606.01247", "kind": "paper", "url": "https://arxiv.org/abs/2606.01247"}, {"id": "gh:aim-uofa/tvrbench", "kind": "repo", "url": "https://github.com/aim-uofa/TVRBench"}], "authors": ["Liyang Li", "Muzhi Zhu", "Zhiyue Zhao", "Hengyu Zhao", "Ke Liu", "Linhao Zhong", "Hao Chen", "Chunhua Shen"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.01247", "source_url": "https://arxiv.org/abs/2606.01247"}, "categories": ["Agents", "Agents & Tool Use", "Geometric reasoning", "Robotics & Embodied AI"], "description": {"en": "A simulated indoor benchmark in which foundation models must actively move and turn to reproduce the viewpoint shown in a target image."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.01247", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.01247", "title": "TVRBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Zhejiang University"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.01247", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.01247", "source_url": "https://openalex.org/W7163218694"}, "key": "claire-radar:2606.01247", "modality": null, "name": "TVRBench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-09-25T04:40:22Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_where-to-look-can-foundation-models-reach-_e17cb55c", "origin_source": "arxiv", "recognition_confidence": "0.7", "record_sha256": "6dfd3f0536be3ee3eb43e7ab39372f458d68cc2153be8b99a903ffc417819713", "relation": "introduces", "review_model": "claude-code-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T16:50:36.102096Z", "source_url": "https://arxiv.org/abs/2606.01247"}, "publisher": null, "released": "2026-05-31", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.01247", "source_url": "https://arxiv.org/abs/2606.01247"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-01247", "source": "claire_radar", "source_benchmark_id": "2606.01247", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Where to Look: Can Foundation Models Reach a Target Viewpoint Through Active Exploration?"], "annotation": "Unknown", "applicationDomains": ["Robotics & Embodied AI"], "area": "Agents & Tool Use", "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "Embodied spatial intelligence is an active topic. The public code, data, and trained models make the benchmark easy to pick up.", "score": 45}, "capabilities": ["Geometric reasoning"], "confidence": "Medium", "construction": "Interactive Environment", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T16:50:36.102096Z", "inputHash": "407eb61b8728e7faf881726f9b1fa3670deb25be8e7a51c024511d2b6249cf99", "model": "claude-code-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "TVRBench, an indoor-simulation benchmark spanning scene scale and target-view visual richness", "canonicalNameSource": "abstract", "decisionReason": "The benchmark is formally named, has a defined evaluation split and a success-rate metric, and its code and data are publicly released.", "evaluationEvidence": {"reuse": "Our code, data, and models are available at https://github.com/aim-uofa/TVRBench.", "scoring": "on the evaluation split, the strongest open-source and closed-source models reach only 7.8% and 12.0% success", "task": "an agent adjusts its viewpoint in a 3D environment until its observation matches a given target image"}, "model": "claude-code-default", "reviewedAt": "2026-09-24T16:50:36.102096Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A simulated indoor benchmark in which foundation models must actively move and turn to reproduce the viewpoint shown in a target image.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "We introduce Target Viewpoint Reproduction (TVR) -- an active task where an agent adjusts its viewpoint in a 3D environment until its observation matches a given target image -- and TVRBench, an indoor-simulation benchmark spanning scene scale and target-view visual richness."}, "familyId": "bmf_d9243524d028", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_where-to-look-can-foundation-models-reach-_e17cb55c", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Robotics"], "links": {"code": "https://github.com/aim-uofa/TVRBench", "data": null, "paper": "https://arxiv.org/abs/2606.01247", "pdf": "https://arxiv.org/pdf/2606.01247", "project": null, "report": "https://arxiv.org/abs/2606.01247"}, "metrics": [], "motivation": "Humans can reproduce the viewpoint specified by a target image through active head and body motion, yet spatial intelligence in foundation models has largely been studied as passive understanding of pre-collected observations.", "name": "TVRBench", "oneLine": "A simulated indoor benchmark in which foundation models must actively move and turn to reproduce the viewpoint shown in a target image.", "paperTitle": "Where to Look: Can Foundation Models Reach a Target Viewpoint Through Active Exploration?", "primaryDomain": "Robotics & Embodied AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.01247", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "publishers": [{"name": "aim-uofa", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/aim-uofa/TVRBench"}], "readiness": "Runnable", "recognitionConfidence": 0.7, "relation": "introduces", "releasedAt": "2026-05-31", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Liyang Li", "Muzhi Zhu", "Zhiyue Zhao", "Hengyu Zhao", "Ke Liu", "Linhao Zhong", "Hao Chen", "Chunhua Shen"], "categories": ["cs.CV"], "id": "2606.01247", "title": "Where to Look: Can Foundation Models Reach a Target Viewpoint Through Active Exploration?", "type": "arxiv", "url": "https://arxiv.org/abs/2606.01247"}, "sourceUpdatedAt": "2026-05-31T14:00:10Z", "topics": ["Agents"], "venueAttempts": [], "whyItMatters": "It tests active, embodied spatial reasoning rather than passive understanding of images. Current models score low on it, which leaves clear room to measure progress."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
