{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2607.04033", "kind": "paper", "url": "https://arxiv.org/abs/2607.04033"}, {"id": "gh:openraiser/omniopt", "kind": "repo", "url": "https://github.com/OpenRaiser/OmniOpt"}], "authors": ["Siyuan Li", "Jiabao Pan", "Yumou Liu", "Zhuoli Ouyang", "Xin Jin", "Xinglong Xu", "Jingxuan Wei", "Shengye Pang", "Jintao Che", "Xuanhe Zhou", "Conghui He", "Cheng Tan"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2607.04033", "source_url": "https://arxiv.org/abs/2607.04033"}, "categories": ["General AI", "Geometric reasoning", "Language & Knowledge", "cs.LG"], "description": {"en": "A cross-domain benchmark for comparing optimizers in large-scale model training. It covers 24+ optimizers across two stages: Stage 1 sweeps on C4 with LLaMA-3 architectures (60M to 1B), and Stage 2 transfers to FineWeb-Edu with four architectures (Transformer++, GLA, DeltaNet, Gated DeltaNet) at 340M and 1B scales. Controlled-variable protocol with fixed architecture, data, and schedule settings."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.04033", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.04033", "title": "OmniOpt"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Westlake University", "Shanghai Artificial Intelligence Laboratory", "Shanghai University", "Shanghai Jiao Tong University", "Southern University of Science and Technology", "University of Chinese Academy of Sciences"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2607.04033", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2607.04033", "source_url": "https://openalex.org/W7167607814"}, "key": "claire-radar:2607.04033", "modality": null, "name": "OmniOpt", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "Medium", "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_omniopt_f648d471", "origin_source": "arxiv", "recognition_confidence": "0.85", "record_sha256": "a931ac3134cc3d5e67af8668e94015593d698a0463921cbb7a913df77db344a0", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2607.04033"}, "publisher": null, "released": "2026-07-04", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.04033", "source_url": "https://arxiv.org/abs/2607.04033"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-04033", "source": "claire_radar", "source_benchmark_id": "2607.04033", "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": "https://github.com/OpenRaiser/OmniOpt", "githubScope": "benchmark_repo", "githubStars": 38, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 54, "hfPaperUrl": "https://huggingface.co/papers/2607.04033", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-04", "state": "stale"}}, "status": "fresh"}, "capabilities": ["Geometric reasoning"], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "297971f39f8cba69d19bc299168f8baf176f30baca420701c21f92ebd23f4510", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A cross-domain benchmark for comparing optimizers in large-scale model training. It covers 24+ optimizers across two stages: Stage 1 sweeps on C4 with LLaMA-3 architectures (60M to 1B), and Stage 2 transfers to FineWeb-Edu with four architectures (Transformer++, GLA, DeltaNet, Gated DeltaNet) at 340M and 1B scales. Controlled-variable protocol with fixed architecture, data, and schedule settings.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "evaluation protocol evidence"], "snippet": "We therefore present OmniOpt, a unified survey and benchmark cookbook of optimizers for the research community."}, "familyId": "bmf_15c40ef3c174", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [{"evidence": [{"basis": "existing-code-link", "linkedUrl": "https://github.com/OpenRaiser/OmniOpt", "url": "https://github.com/OpenRaiser/OmniOpt"}, {"basis": "existing-metric-association", "linkedUrl": "https://github.com/OpenRaiser/OmniOpt", "url": "https://huggingface.co/papers/2607.04033"}], "scope": "benchmark_repo", "url": "https://github.com/OpenRaiser/OmniOpt", "verification": "source-linked"}], "status": "linked"}, "heat": null, "id": "bm_omniopt_f648d471", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/OpenRaiser/OmniOpt", "data": null, "hfPaper": "https://huggingface.co/papers/2607.04033", "paper": "https://arxiv.org/abs/2607.04033", "pdf": "https://arxiv.org/pdf/2607.04033", "project": null, "report": "https://arxiv.org/abs/2607.04033"}, "metrics": [], "motivation": "Optimizer selection for large-scale model training has become a system-level design decision constrained jointly by compute, memory, tuning budget, and task diversity, yet the landscape of over one hundred methods remains fragmented.", "name": "OmniOpt", "oneLine": "A cross-domain benchmark for comparing optimizers in large-scale model training. It covers 24+ optimizers across two stages: Stage 1 sweeps on C4 with LLaMA-3 architectures (60M to 1B), and Stage 2 transfers to FineWeb-Edu with four architectures (Transformer++, GLA, DeltaNet, Gated DeltaNet) at 340M and 1B scales. Controlled-variable protocol with fixed architecture, data, and schedule settings.", "paperTitle": "OmniOpt: Taxonomy, Geometry, and Benchmarking of Modern Optimizers", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2607.04033", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "OmniOpt Team", "organizationType": "community", "role": "benchmark-publisher", "sourceUrl": "https://github.com/OpenRaiser/OmniOpt"}], "ranking": {}, "readiness": "Runnable", "recognitionConfidence": 0.85, "relation": "introduces", "releasedAt": "2026-07-04", "source": {"authors": ["Siyuan Li", "Jiabao Pan", "Yumou Liu", "Zhuoli Ouyang", "Xin Jin", "Xinglong Xu", "Jingxuan Wei", "Shengye Pang", "Jintao Che", "Xuanhe Zhou", "Conghui He", "Cheng Tan"], "categories": ["cs.LG", "cs.AI"], "id": "2607.04033", "title": "OmniOpt: Taxonomy, Geometry, and Benchmarking of Modern Optimizers", "type": "arxiv", "url": "https://arxiv.org/abs/2607.04033"}, "sourceUpdatedAt": "2026-07-04T21:27:05Z", "topics": ["cs.LG"], "venueAttempts": [], "whyItMatters": "Optimizer selection is a system-level decision impacting compute, memory, and tuning budget. This benchmark provides a unified protocol for comparing methods across multiple scales and architectures, offering reproducible evidence for practitioners to choose optimizers based on measured training objectives and trade-offs."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
