{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.07682", "kind": "paper", "url": "https://arxiv.org/abs/2606.07682"}, {"kind": "website", "url": "https://swe-marathon.org/"}], "authors": ["Rishi Desai", "Jesse Hu", "Joan Cabezas", "Neel Harsola", "Pratyush Shukla", "Roey Ben Chaim", "Adnan El Assadi", "Omkaar Mukund Kamath", "Fenil Faldu", "Prannay Hebbar", "Jiankai Sun", "Yiyuan Li", "Pramod Srinivasan", "Ishan Gupta", "Christopher Settles", "Daniel Wang", "Derek Chen", "Pranav Raja", "Albert Liu", "Marek Šuppa", "Nevasini Sasikumar", "Luyang Kong", "Erik Quintanilla", "Xiangyi Li", "Ivan Bercovich", "Steven Dillmann"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.07682", "source_url": "https://arxiv.org/abs/2606.07682"}, "categories": ["Code", "Code & Software", "Software & AI Compute"], "description": {"en": "SWE-Marathon evaluates AI agents on 20 ultra-long-horizon software engineering tasks, each with a unique executable environment, a human-written reference solution, and a multi-layer verification suite. Tasks average 27.2M tokens per logged agent attempt, requiring sustained progress over hours and millions of tokens."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.07682", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.07682", "title": "SWE-Marathon"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Gujarat Technological University", "Harvard University", "University of Waterloo", "University of North Carolina at Chapel Hill", "Stanford University", "Georgia Institute of Technology", "Comenius University Bratislava", "University of California, Santa Cruz"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.07682", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.07682", "source_url": "https://openalex.org/W7164012646"}, "key": "claire-radar:2606.07682", "modality": null, "name": "SWE-Marathon", "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_swe-marathon_2e8e782b", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "853e6f3acd17c139b9b525df43192d5ad434598871cd3a75b787023a2b3bf89b", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.07682"}, "publisher": null, "released": "2026-06-05", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.07682", "source_url": "https://arxiv.org/abs/2606.07682"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-07682", "source": "claire_radar", "source_benchmark_id": "2606.07682", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["Software & AI Compute"], "area": "Code & Software", "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": 2, "hfPaperUrl": "https://huggingface.co/papers/2606.07682", "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", "lastSuccessfulDate": "2026-09-25", "state": "stale"}}, "status": "stale"}, "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": "4a8516fea7e8c588b9b3aa7d8155f8d86deddf6e4bbc0af148466d0996221ea2", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "SWE-Marathon evaluates AI agents on 20 ultra-long-horizon software engineering tasks, each with a unique executable environment, a human-written reference solution, and a multi-layer verification suite. Tasks average 27.2M tokens per logged agent attempt, requiring sustained progress over hours and millions of tokens.", "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 SWE-Marathon, a benchmark of 20 long-horizon tasks spanning software engineering and adjacent technical domains."}, "familyId": "bmf_fbbe5c4b7d4a", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_swe-marathon_2e8e782b", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Software & Cloud"], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2606.07682", "paper": "https://arxiv.org/abs/2606.07682", "pdf": "https://arxiv.org/pdf/2606.07682", "project": "https://swe-marathon.org/", "report": "https://arxiv.org/abs/2606.07682"}, "metrics": [], "motivation": "AI agents are increasingly expected to complete long-horizon workflows that require sustained progress over hours, millions of tokens, and complex environments.", "name": "SWE-Marathon", "oneLine": "SWE-Marathon evaluates AI agents on 20 ultra-long-horizon software engineering tasks, each with a unique executable environment, a human-written reference solution, and a multi-layer verification suite. Tasks average 27.2M tokens per logged agent attempt, requiring sustained progress over hours and millions of tokens.", "paperTitle": "SWE-Marathon: Can Agents Autonomously Complete Ultra-Long-Horizon Software Work?", "primaryDomain": "Software & AI Compute", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.07682", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "swe-marathon.org", "organizationType": "community", "role": "benchmark-publisher", "sourceUrl": "https://swe-marathon.org/"}], "ranking": {}, "readiness": "Inspectable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-06-05", "source": {"authors": ["Rishi Desai", "Jesse Hu", "Joan Cabezas", "Neel Harsola", "Pratyush Shukla", "Roey Ben Chaim", "Adnan El Assadi", "Omkaar Mukund Kamath", "Fenil Faldu", "Prannay Hebbar", "Jiankai Sun", "Yiyuan Li", "Pramod Srinivasan", "Ishan Gupta", "Christopher Settles", "Daniel Wang", "Derek Chen", "Pranav Raja", "Albert Liu", "Marek \\v{S}uppa", "Nevasini Sasikumar", "Luyang Kong", "Erik Quintanilla", "Xiangyi Li", "Ivan Bercovich", "Steven Dillmann"], "categories": ["cs.SE", "cs.AI"], "id": "2606.07682", "title": "SWE-Marathon: Can Agents Autonomously Complete Ultra-Long-Horizon Software Work?", "type": "arxiv", "url": "https://arxiv.org/abs/2606.07682"}, "sourceUpdatedAt": "2026-06-05T00:39:44Z", "topics": ["Code"], "venueAttempts": [], "whyItMatters": "Existing agent benchmarks focus on short tasks, limiting measurement of planning, long-context understanding, and memory. SWE-Marathon addresses the gap by providing a longer-horizon evaluation that exposes practical limitations in agent autonomy and highlights failure modes like reward hacking, informing development of more robust agents."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
