{"record": {"aliases": ["MM-ToolSandBox: A Unified Framework for Evaluating Visual Tool-Calling Agents"], "artifacts": [{"id": "arxiv:2607.11818", "kind": "paper", "url": "https://arxiv.org/abs/2607.11818"}, {"id": "gh:apple/ml-mmtoolsandbox", "kind": "repo", "url": "https://github.com/apple/ml-mmtoolsandbox"}], "categories": ["General AI", "Vision & 3D", "cs.CV"], "description": {"en": "MM-ToolSandBox is a stateful sandbox benchmark for visually grounded tool-calling agents. It has 258 human-verified scenarios across 500+ tools in 16 domains, and models are scored by task success rate."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.11818", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.11818", "title": "MM-ToolSandBox"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2607.11818", "modality": null, "name": "MM-ToolSandBox", "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_mm-toolsandbox-a-unified-framework-for-eva_3d2f4872", "origin_source": "arxiv", "recognition_confidence": "0.55", "record_sha256": "972c4b84e0e603af2e0ec7072bd91cf0515b3a98f7c1b7c5172af96ac963f972", "relation": "evaluates_only", "review_model": "claude-code-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T20:39:58.972873Z", "source_url": "https://arxiv.org/abs/2607.11818"}, "publisher": null, "released": "2026-07-13", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.11818", "source_url": "https://arxiv.org/abs/2607.11818"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-11818", "source": "claire_radar", "source_benchmark_id": "2607.11818", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["MM-ToolSandBox: A Unified Framework for Evaluating Visual Tool-Calling Agents"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Vision & 3D", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The public repository and the evaluation of 12 models make it easy to adopt, and agent tool-use evaluation is an active area.", "score": 62}, "capabilities": [], "confidence": "Medium", "construction": "Transform Existing", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T20:39:58.972873Z", "inputHash": "4bc3aaac5a426aaa988bbc96689dca7c31ab500d471d7939b8d29ad0bb58d5f1", "model": "claude-code-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "MM-ToolSandBox: A Unified Framework for Evaluating Visual Tool-Calling Agents", "canonicalNameSource": "paper_title", "decisionReason": "The benchmark is formally named, has a defined task environment, uses success rate as a comparable metric, and is publicly available through a GitHub repository.", "evaluationEvidence": {"reuse": "The framework and the benchmark are publicly available at https://github.com/apple/ml-mmtoolsandbox", "scoring": "even the best model achieves below 50% success rate", "task": "a stateful execution environment spanning 500+ tools across 16 application domains, supporting multi-image, multi-turn tasks"}, "model": "claude-code-default", "reviewedAt": "2026-09-24T20:39:58.972873Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "MM-ToolSandBox is a stateful sandbox benchmark for visually grounded tool-calling agents. It has 258 human-verified scenarios across 500+ tools in 16 domains, and models are scored by task success rate.", "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", "named entity is a model, method, framework, or dataset used in evaluation"], "snippet": "We introduce MM-ToolSandBox, a benchmark and evaluation framework for visually grounded tool-calling agents."}, "familyId": "bmf_f4c3782439f5", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_mm-toolsandbox-a-unified-framework-for-eva_3d2f4872", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/apple/ml-mmtoolsandbox", "data": null, "paper": "https://arxiv.org/abs/2607.11818", "pdf": "https://arxiv.org/pdf/2607.11818", "project": null, "report": "https://arxiv.org/abs/2607.11818"}, "metrics": [], "motivation": "We introduce MM-ToolSandBox, a benchmark and evaluation framework for visually grounded tool-calling agents.", "name": "MM-ToolSandBox", "oneLine": "MM-ToolSandBox is a stateful sandbox benchmark for visually grounded tool-calling agents. It has 258 human-verified scenarios across 500+ tools in 16 domains, and models are scored by task success rate.", "paperTitle": "MM-ToolSandBox: A Unified Framework for Evaluating Visual Tool-Calling Agents", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2607.11818", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "publishers": [{"name": "Apple", "organizationType": "company-research-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/apple/ml-mmtoolsandbox"}], "readiness": "Runnable", "recognitionConfidence": 0.55, "relation": "evaluates_only", "releasedAt": "2026-07-13", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Kaixin Ma", "Di Feng", "Alexander Metz", "Jiarui Lu", "Eshan Verma", "Afshin Dehghan"], "categories": ["cs.CV", "cs.AI"], "id": "2607.11818", "title": "MM-ToolSandBox: A Unified Framework for Evaluating Visual Tool-Calling Agents", "type": "arxiv", "url": "https://arxiv.org/abs/2607.11818"}, "sourceUpdatedAt": "2026-07-13T17:13:09Z", "topics": ["cs.CV"], "venueAttempts": [], "whyItMatters": "It targets multi-image, multi-turn visual tool use and separates perception failures from planning failures. The paper reports that incorrect information extraction from images causes 53% of failures."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
