{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2605.24110", "kind": "paper", "url": "https://arxiv.org/abs/2605.24110"}], "authors": ["Haiyang Shen", "Xuanzhong Chen", "Wendong Xu", "Yun Ma", "Liang Chen", "Kuan Li"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2605.24110", "source_url": "https://arxiv.org/abs/2605.24110"}, "categories": ["General AI", "Language & Knowledge", "cs.AI"], "description": {"en": "EvoCode-Bench evaluates coding agents in multi-turn iterative interactions, with 26 stateful coding tasks and 227 evaluated rounds. Each task preserves the agent's workspace for 5-15 rounds, specifies requirements via observable behavior, and uses cumulative executable tests to verify new and prior requirements. Scoring uses MT@4, a four-attempt fail-stop multi-round score, and SR, a single-round score from a reference-completed prior state."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2605.24110", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2605.24110", "title": "EvoCode-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Peking University", "Tsinghua University", "University of Hong Kong"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2605.24110", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2605.24110", "source_url": "https://openalex.org/W7162397884"}, "key": "claire-radar:2605.24110", "modality": null, "name": "EvoCode-Bench", "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_evocode-bench_60a115e8", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "7e3f4f239a487352beb761fca54fb114caf73f5bfc03cee703bdee06eea65826", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2605.24110"}, "publisher": null, "released": "2026-05-22", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2605.24110", "source_url": "https://arxiv.org/abs/2605.24110"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2605-24110", "source": "claire_radar", "source_benchmark_id": "2605.24110", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Mixed", "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": 1, "hfPaperUrl": "https://huggingface.co/papers/2605.24110", "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": "Aggregate Existing", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "6c10356ec27b25e4d29c0d964ed27fd97db09167b3f369ba416cb2b21b39da24", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "EvoCode-Bench evaluates coding agents in multi-turn iterative interactions, with 26 stateful coding tasks and 227 evaluated rounds. Each task preserves the agent's workspace for 5-15 rounds, specifies requirements via observable behavior, and uses cumulative executable tests to verify new and prior requirements. Scoring uses MT@4, a four-attempt fail-stop multi-round score, and SR, a single-round score from a reference-completed prior state.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We introduce EvoCode-Bench, a benchmark of 26 stateful coding tasks and 227 evaluated rounds."}, "familyId": "bmf_7accae8d1fc6", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_evocode-bench_60a115e8", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2605.24110", "paper": "https://arxiv.org/abs/2605.24110", "pdf": "https://arxiv.org/pdf/2605.24110", "project": null, "report": "https://arxiv.org/abs/2605.24110"}, "metrics": [], "motivation": "Coding agents are increasingly used as iterative development partners, but most benchmarks still evaluate one specification followed by one final assessment.", "name": "EvoCode-Bench", "oneLine": "EvoCode-Bench evaluates coding agents in multi-turn iterative interactions, with 26 stateful coding tasks and 227 evaluated rounds. Each task preserves the agent's workspace for 5-15 rounds, specifies requirements via observable behavior, and uses cumulative executable tests to verify new and prior requirements. Scoring uses MT@4, a four-attempt fail-stop multi-round score, and SR, a single-round score from a reference-completed prior state.", "paperTitle": "EvoCode-Bench: Evaluating Coding Agents in Multi-Turn Iterative Interactions", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2605.24110", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "Harbor", "organizationType": "community", "role": "benchmark-publisher", "sourceUrl": "https://huggingface.co/papers/2605.24110"}], "ranking": {}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-05-22", "source": {"authors": ["Haiyang Shen", "Xuanzhong Chen", "Wendong Xu", "Yun Ma", "Liang Chen", "and Kuan Li"], "categories": ["cs.AI"], "id": "2605.24110", "title": "EvoCode-Bench: Evaluating Coding Agents in Multi-Turn Iterative Interactions", "type": "arxiv", "url": "https://arxiv.org/abs/2605.24110"}, "sourceUpdatedAt": "2026-05-22T18:17:28Z", "topics": ["cs.AI"], "venueAttempts": [], "whyItMatters": "Existing coding benchmarks typically evaluate a single specification and final output, missing the ability to handle evolving requirements. EvoCode-Bench fills this gap by assessing whether agents can maintain a working codebase over multiple rounds. The gap between SR and MT@4 scores reveals that high single-turn performance does not guarantee sustained multi-turn success, providing a more relevant evaluation for real-world iterative development."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
