{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.20318", "kind": "paper", "url": "https://arxiv.org/abs/2608.20318"}], "authors": ["Yizhe Chi", "Wenyi Li", "Deyao Hong", "Xiaoqiu Wang", "Mingju Gao", "Kaisen Yang", "Bingxiang He", "Youjie Zheng", "Calvin Xiao", "Qinhuai Na"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.20318", "source_url": "https://arxiv.org/abs/2608.20318"}, "categories": ["Language & Knowledge", "Self-Evolution", "Software & AI Compute"], "description": {"en": "AI4AI-Bench evaluates LLM agents on algorithmic design across 10 frozen research repositories. Agents rewrite training algorithms within 4 hours, then scored by fixed evaluators against baseline algorithms, with submissions released for repeatable measurement."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.20318", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.20318", "title": "AI4AI-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Tsinghua University"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2608.20318", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2608.20318", "source_url": "https://openalex.org/W7203900641"}, "key": "claire-radar:2608.20318", "modality": null, "name": "AI4AI-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": "true", "origin_record_id": "bm_ai4ai-bench_cbcf9954", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "be55b28088d8c4a8cdf403c4ef397cee51067cefd8342e121041de0071d5cfd5", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-08-25T15:38:21.596820Z", "source_url": "https://arxiv.org/abs/2608.20318"}, "publisher": null, "released": "2026-08-20", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.20318", "source_url": "https://arxiv.org/abs/2608.20318"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-20318", "source": "claire_radar", "source_benchmark_id": "2608.20318", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["Software & AI Compute"], "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": 3, "hfPaperUrl": "https://huggingface.co/papers/2608.20318", "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": "2026-09-30", "state": "stale"}}, "status": "stale"}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-25T16:23:14.196185Z", "inputHash": "118254959b20a36622cdf7b21728a8c49bc197748e70a39485e6b057966e2d0a", "model": "deepseek-v4-flash"}, "curation": {"canonicalNameEvidence": "AI4AI-Bench: Benchmarking LLM Agents in Algorithmic Design for Recursive Self-Improvement", "canonicalNameSource": "paper_title", "decisionReason": "Formally named benchmark with released task suite, evaluators, and scored submissions, enabling repeated measurement.", "model": "deepseek-v4-pro", "reviewedAt": "2026-08-25T15:38:21.596820Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "AI4AI-Bench evaluates LLM agents on algorithmic design across 10 frozen research repositories. Agents rewrite training algorithms within 4 hours, then scored by fixed evaluators against baseline algorithms, with submissions released for repeatable measurement.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We present AI4AI\\mbox{-}Bench, 10 frozen research repositories spanning 10 training algorithm families."}, "familyId": "bmf_300170ad25dc", "firstSeenAt": "2026-08-23", "heat": null, "id": "bm_ai4ai-bench_cbcf9954", "indexedAt": "2026-08-23T06:55:50.392548Z", "industrySectors": ["Software & Cloud"], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2608.20318", "pdf": "https://arxiv.org/pdf/2608.20318", "project": null, "report": "https://arxiv.org/abs/2608.20318"}, "metrics": [], "motivation": "Recursive self-improvement (RSI) asks whether an AI system can improve the process that produces AI systems, so that the next system inherits the improvement.", "name": "AI4AI-Bench", "oneLine": "AI4AI-Bench evaluates LLM agents on algorithmic design across 10 frozen research repositories. Agents rewrite training algorithms within 4 hours, then scored by fixed evaluators against baseline algorithms, with submissions released for repeatable measurement.", "paperTitle": "AI4AI-Bench: Benchmarking LLM Agents in Algorithmic Design for Recursive Self-Improvement", "primaryDomain": "Software & AI Compute", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.20318", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-23T06:55:50.392548Z"}, "publications": [], "publishers": [{"name": "AI4AI-Bench team", "organizationType": "benchmark-organization", "role": "benchmark-publisher", "sourceUrl": "https://arxiv.org/abs/2608.20318"}], "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": 0.24495074076910422, "value": 3}}, "confidence": "Low", "coverage": 0.15, "rank": null, "score": 46}, "method": "attention-ranking-v13", "rank": null, "score": 46}}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-08-20", "source": {"authors": ["Yizhe Chi", "Wenyi Li", "Deyao Hong", "Xiaoqiu Wang", "Mingju Gao", "Kaisen Yang", "Bingxiang He", "Youjie Zheng", "Calvin Xiao", "Qinhuai Na"], "categories": ["cs.AI", "cs.CL", "cs.LG"], "id": "2608.20318", "title": "AI4AI-Bench: Benchmarking LLM Agents in Algorithmic Design for Recursive Self-Improvement", "type": "arxiv", "url": "https://arxiv.org/abs/2608.20318"}, "sourceUpdatedAt": "2026-08-20T17:56:59Z", "topics": ["Self-Evolution"], "venueAttempts": [], "whyItMatters": "It isolates the ability to design training algorithms for recursive self-improvement, a capability not directly measured by existing benchmarks, and provides a concrete, normalized scoring scale."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
