{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2607.10079", "kind": "paper", "url": "https://arxiv.org/abs/2607.10079"}], "authors": ["Chengguang Gan", "Hanjun Wei", "Yunhao Liang", "Zhixi Cai", "Qinghao Zhang", "Shiwen Ni"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2607.10079", "source_url": "https://arxiv.org/abs/2607.10079"}, "categories": ["Agents", "General AI", "Multimodal"], "description": {"en": "MAG is a benchmark that unifies task execution and guide writing into a single multimodal action and guide task, with grounding over screenshots. It includes a harness for annotation, training, evaluation, and joint metrics."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.10079", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.10079", "title": "MAG"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["University of Chinese Academy of Sciences", "Monash University", "Pusan National University", "Shenzhen University of Advanced Technology"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2607.10079", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2607.10079", "source_url": "https://openalex.org/W7168238693"}, "key": "claire-radar:2607.10079", "modality": null, "name": "MAG", "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_mag_3c2b6829", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "8b44bb45e1506cd054bc5d43ef29edbf38576a4be56c82ae4e48a380db043074", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2607.10079"}, "publisher": null, "released": "2026-07-11", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.10079", "source_url": "https://arxiv.org/abs/2607.10079"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-10079", "source": "claire_radar", "source_benchmark_id": "2607.10079", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Multimodal", "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": 0, "hfPaperUrl": "https://huggingface.co/papers/2607.10079", "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": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "f07973bc7589c7ceb062e24adbecc848593c2bb8b8b2d41f9e3336833c1996ce", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "MAG is a benchmark that unifies task execution and guide writing into a single multimodal action and guide task, with grounding over screenshots. It includes a harness for annotation, training, evaluation, and joint metrics.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "In this work we introduce MAG, the first benchmark that unifies task execution and guide writing into a single Multimodal Action and Guide task, with two grounding schemes over screenshots: Set-of-Mark element selection and raw pixel coordinates."}, "familyId": "bmf_f8685538bda2", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_mag_3c2b6829", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2607.10079", "paper": "https://arxiv.org/abs/2607.10079", "pdf": "https://arxiv.org/pdf/2607.10079", "project": null, "report": "https://arxiv.org/abs/2607.10079"}, "metrics": [], "motivation": "Digital Adoption Platforms (DAPs) are embedded overlays widely used on web systems to guide users through operations inside a page, helping them get started with unfamiliar interfaces quickly.", "name": "MAG", "oneLine": "MAG is a benchmark that unifies task execution and guide writing into a single multimodal action and guide task, with grounding over screenshots. It includes a harness for annotation, training, evaluation, and joint metrics.", "paperTitle": "MAG: A Web-Agent Benchmark and Harness for Multimodal Action and Guide Generation", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2607.10079", "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": 0.0, "value": 0}}, "confidence": "Low", "coverage": 0.15, "rank": null, "score": 43}, "method": "attention-ranking-v13", "rank": null, "score": 43}}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-07-11", "source": {"authors": ["Chengguang Gan", "Hanjun Wei", "Yunhao Liang", "Zhixi Cai", "Qinghao Zhang", "Shiwen Ni"], "categories": ["cs.AI", "cs.CL"], "id": "2607.10079", "title": "MAG: A Web-Agent Benchmark and Harness for Multimodal Action and Guide Generation", "type": "arxiv", "url": "https://arxiv.org/abs/2607.10079"}, "sourceUpdatedAt": "2026-07-16T03:28:43Z", "topics": ["Agents", "Multimodal"], "venueAttempts": [], "whyItMatters": "The evaluation gap is that prior benchmarks separate web agent actions and guide text generation, and often rely on textual DOM rather than screenshots. MAG provides a unified evaluation for multimodal understanding and generation in live environments, which could support development of more capable web agents."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
