{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.08392", "kind": "paper", "url": "https://arxiv.org/abs/2608.08392"}, {"kind": "website", "url": "https://warriorxu0302.github.io/CAP-Bench/"}], "authors": ["Zejun Xu", "Taiyi Chen", "Jin Li", "Yongtong Gu", "Qi Cheng", "Aixuan Lv", "Shuai Zhu", "Pengfei Zhu", "Kaichen Yang", "Boyu Sun", "Yixian Yang", "Mulong Xie", "Xin Liu", "Dagang Li", "Xiaoteng Ma", "Hongru Wang"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.08392", "source_url": "https://arxiv.org/abs/2608.08392"}, "categories": ["General AI", "Language & Knowledge", "cs.AI"], "description": {"en": "CAP evaluates cross-site browser agents on 420 tasks across 108 real websites and 24 domains. Tasks require complex UI interactions and visual perception, with scoring via an agent-as-a-judge framework."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.08392", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.08392", "title": "CAP"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Macau University of Science and Technology", "Tsinghua University", "Southeast University", "University of Edinburgh"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2608.08392", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2608.08392", "source_url": "https://openalex.org/W7202205692"}, "key": "claire-radar:2608.08392", "modality": null, "name": "CAP", "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_cap_c96b1f01", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "ebd8b4b82b3c00dfac600321ba1cebcd4bf42fa60b34a83d4b1bbecc06c1222d", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2608.08392"}, "publisher": null, "released": "2026-08-09", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.08392", "source_url": "https://arxiv.org/abs/2608.08392"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-08392", "source": "claire_radar", "source_benchmark_id": "2608.08392", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "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": null, "hfPaperUrl": "https://huggingface.co/papers/2608.08392", "observedAt": "2026-10-04T03:02:42.754940Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "state": "unavailable"}}, "status": "unavailable"}, "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": "2477e8d117b2c7ccb623acd131c8c26a2ed1022dbd26a5b62590a12880799d86", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "CAP evaluates cross-site browser agents on 420 tasks across 108 real websites and 24 domains. Tasks require complex UI interactions and visual perception, with scoring via an agent-as-a-judge framework.", "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", "public artifact URL"], "snippet": "We introduce CAP, a scalable benchmark for evaluating browser agents on cross-site, human-like web tasks that require non-trivial UI interactions and visual understanding."}, "familyId": "bmf_45e84635bc48", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 1, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "candidates"}, "heat": null, "id": "bm_cap_c96b1f01", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2608.08392", "paper": "https://arxiv.org/abs/2608.08392", "pdf": "https://arxiv.org/pdf/2608.08392", "project": "https://warriorxu0302.github.io/CAP-Bench/", "report": "https://arxiv.org/abs/2608.08392"}, "metrics": [], "motivation": "Large language models are increasingly deployed as autonomous agents that interact with the web through browsers.", "name": "CAP", "oneLine": "CAP evaluates cross-site browser agents on 420 tasks across 108 real websites and 24 domains. Tasks require complex UI interactions and visual perception, with scoring via an agent-as-a-judge framework.", "paperTitle": "CAP: A Scalable Benchmark for Evaluating Cross-Site Browser Agents with Complex Actions and Perception", "primaryDomain": "General AI", "publication": {"evidence": "Accepted to COLM 2026. Project page: https://warriorxu0302.github.io/CAP-Bench/", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2608.08392", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "COLM 2026", "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "CAP-Bench Project", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://warriorxu0302.github.io/CAP-Bench/"}], "ranking": {"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": "Inspectable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-08-09", "source": {"authors": ["Zejun Xu", "Taiyi Chen", "Jin Li", "Yongtong Gu", "Qi Cheng", "Aixuan Lv", "Shuai Zhu", "Pengfei Zhu", "Kaichen Yang", "Boyu Sun", "Yixian Yang", "Mulong Xie", "Xin Liu", "Dagang Li", "Xiaoteng Ma", "and Hongru Wang"], "categories": ["cs.AI", "cs.CL"], "id": "2608.08392", "title": "CAP: A Scalable Benchmark for Evaluating Cross-Site Browser Agents with Complex Actions and Perception", "type": "arxiv", "url": "https://arxiv.org/abs/2608.08392"}, "sourceUpdatedAt": "2026-08-09T01:08:19Z", "topics": ["cs.AI"], "venueAttempts": [{"decisionRaw": "Accepted to COLM 2026. Project page: https://warriorxu0302.github.io/CAP-Bench/", "evidence": [{"level": "author-claim", "observedAt": "2026-10-14T01:46:58.058199Z", "rawValue": "Accepted to COLM 2026. Project page: https://warriorxu0302.github.io/CAP-Bench/", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2608.08392"}], "reviewStatus": "accepted", "venueName": "COLM 2026"}], "whyItMatters": "CAP addresses the gap in browser agent evaluation by focusing on cross-site workflows and perception-heavy interactions, which are common in real-world browsing. It provides fine-grained diagnostics to identify bottlenecks in current agents, aiding targeted improvements."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
