{ "profile": "wrs-published-research-view-1", "edition": "ratings-e039106f30dec177", "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS", "entry": { "id": "liu-et-al-moonlight-scalable-muon-2025", "title": "Moonlight: update scaling and distributed Muon training", "summary": "Added weight decay and adjusted parameter-update scales for large-model Muon, implementing ZeRO-style distributed orthogonalization and releasing Moonlight checkpoints with scaling experiments.", "boundary": "Extends rather than invents Muon/NewtonSchulz/ZeRO, and differs from KimiK2 MuonClip. Approximately twofold compute efficiency is fitted/reported for studied training regimes, not universal wall-clock superiority or tuning-free performance for every architecture.", "area": "Large-language-model optimization systems", "topic": "Artificial intelligence", "year": 2025, "lane": "HUMAN", "result": "ALGORITHM", "package_id": "liu-et-al-moonlight-scalable-muon-2025", "target": { "record_id": "urn:uuid:5fada9ce-dc6c-482d-a63e-2b9921a89d3c", "version": 1, "content_hash": "sha256:87ee892e0bc4f7c9a26c59ecf5bbec4bf1ca6d90fabfb48906fdf61c94b56463" }, "humans": [ "Jingyuan Liu", "Jianlin Su", "Xingcheng Yao", "Zhejun Jiang", "Guokun Lai", "Yulun Du", "Yidao Qin", "Weixin Xu", "Enzhe Lu", "Junjie Yan", "Yanru Chen", "Huabin Zheng", "Yibo Liu", "Shaowei Liu", "Bohong Yin", "Weiran He", "Han Zhu", "Yuzhi Wang", "Jianzhou Wang", "Mengnan Dong", "Zheng Zhang", "Yongsheng Kang", "Hao Zhang", "Xinran Xu", "Yutao Zhang", "Yuxin Wu", "Xinyu Zhou", "Zhilin Yang" ], "authors_complete": true, "human_role_note": "Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.", "agents": [], "related": [], "subresults": [], "scientific_assessment": { "source_review": "SOURCE_CURATED", "scientific_review": "NOT_REVIEWED_BY_WRS", "formal_replay": "NOT_RUN", "provider_attestation": "NONE" }, "sources": [ { "id": "cur20261015-arxiv-2502-16982v1", "title": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } ], "points": "70000", "published_grade": { "id": "liu-et-al-moonlight-scalable-muon-2025", "entry_id": "liu-et-al-moonlight-scalable-muon-2025", "band": "B", "modifier": "+", "target": { "record_id": "urn:uuid:5fada9ce-dc6c-482d-a63e-2b9921a89d3c", "version": 1, "content_hash": "sha256:87ee892e0bc4f7c9a26c59ecf5bbec4bf1ca6d90fabfb48906fdf61c94b56463" }, "what_changed": "Added weight decay and adjusted parameter-update scales for large-model Muon, implementing ZeRO-style distributed orthogonalization and releasing Moonlight checkpoints with scaling experiments.", "scope": "Added weight decay and adjusted parameter-update scales for large-model Muon, implementing ZeRO-style distributed orthogonalization and releasing Moonlight checkpoints with scaling experiments. Extends rather than invents Muon/NewtonSchulz/ZeRO, and differs from KimiK2 MuonClip. Approximately twofold compute efficiency is fitted/reported for studied training regimes, not universal wall-clock superiority or tuning-free performance for every architecture.", "status": "PROVISIONAL_EDITORIAL", "scientific_correctness_certified": false }, "assessment_disposition": { "status": "ASSESSED", "label": "Provisionally assessed", "reason": "A current grade and separate reference score identify this exact contribution.", "target": { "record_id": "urn:uuid:5fada9ce-dc6c-482d-a63e-2b9921a89d3c", "version": 1, "content_hash": "sha256:87ee892e0bc4f7c9a26c59ecf5bbec4bf1ca6d90fabfb48906fdf61c94b56463" }, "edition": "ratings-e039106f30dec177" } }, "version": { "kind": "PUBLIC_RECORD_VIEW", "previous": null, "published_at": "2026-10-16T00:00:00Z", "publisher": { "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33", "role": "WRS_EDITORIAL_PUBLICATION" }, "record": { "contributors": [ { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Jingyuan Liu", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Jianlin Su", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Xingcheng Yao", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Zhejun Jiang", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Guokun Lai", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Yulun Du", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Yidao Qin", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Weixin Xu", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Enzhe Lu", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Junjie Yan", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Yanru Chen", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Huabin Zheng", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Yibo Liu", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Shaowei Liu", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Bohong Yin", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Weiran He", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Han Zhu", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Yuzhi Wang", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Jianzhou Wang", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Mengnan Dong", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Zheng Zhang", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Yongsheng Kang", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Hao Zhang", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Xinran Xu", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Yutao Zhang", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Yuxin Wu", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Xinyu Zhou", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Zhilin Yang", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } } ], "evidence": [ { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-0", "kind": "source_reference", "source": { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } } ], "origin": "HUMAN_LEGACY", "record_profile": "wrs-local-research-notes-0.1", "record_type": "research_note", "rrp_version": "0.1", "scope": { "description": "Extends rather than invents Muon/NewtonSchulz/ZeRO, and differs from KimiK2 MuonClip. Approximately twofold compute efficiency is fitted/reported for studied training regimes, not universal wall-clock superiority or tuning-free performance for every architecture.", "domain": "Large-language-model optimization systems" }, "sources": [ { "citation": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1" } ], "statement": "# Moonlight: update scaling and distributed Muon training\n\nAdded weight decay and adjusted parameter-update scales for large-model Muon, implementing ZeRO-style distributed orthogonalization and releasing Moonlight checkpoints with scaling experiments.\n\n## Scientific scope\n\nExtends rather than invents Muon/NewtonSchulz/ZeRO, and differs from KimiK2 MuonClip. Approximately twofold compute efficiency is fitted/reported for studied training regimes, not universal wall-clock superiority or tuning-free performance for every architecture.\n\n## Reading status\n\nThis is a source-linked WRS catalogue note. A full explanatory guide and independent scientific reconstruction have not been completed for this entry. Read the identified original work for its complete narrative and arguments.\n\n## People and systems\n\nJingyuan Liu; Jianlin Su; Xingcheng Yao; Zhejun Jiang; Guokun Lai; Yulun Du; Yidao Qin; Weixin Xu; Enzhe Lu; Junjie Yan; Yanru Chen; Huabin Zheng; Yibo Liu; Shaowei Liu; Bohong Yin; Weiran He; Han Zhu; Yuzhi Wang; Jianzhou Wang; Mengnan Dong; Zheng Zhang; Yongsheng Kang; Hao Zhang; Xinran Xu; Yutao Zhang; Yuxin Wu; Xinyu Zhou; Zhilin Yang\n\nSource-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.\n\n\n\n## Original sources\n\n- [Muon is Scalable for LLM Training](https://arxiv.org/abs/2502.16982v1)\n", "summary": "Added weight decay and adjusted parameter-update scales for large-model Muon, implementing ZeRO-style distributed orthogonalization and releasing Moonlight checkpoints with scaling experiments.", "title": "Moonlight: update scaling and distributed Muon training" }, "record_id": "urn:uuid:5fada9ce-dc6c-482d-a63e-2b9921a89d3c", "version": 1, "source_ref": { "record_id": "urn:uuid:5fada9ce-dc6c-482d-a63e-2b9921a89d3c", "version": 1, "content_hash": "sha256:87ee892e0bc4f7c9a26c59ecf5bbec4bf1ca6d90fabfb48906fdf61c94b56463" }, "export_view": "wrs-public-editorial-view-1", "public_view_hash": "sha256:6e872104f051da4932c81a3b8a5114b1990f630ed02727b059f97e6a1d37111b" }, "target": { "record_id": "urn:uuid:5fada9ce-dc6c-482d-a63e-2b9921a89d3c", "version": 1, "content_hash": "sha256:87ee892e0bc4f7c9a26c59ecf5bbec4bf1ca6d90fabfb48906fdf61c94b56463" }, "published_grade": { "profile": "wrs-published-grade-1", "edition": "2026-10-28.1", "id": "liu-et-al-moonlight-scalable-muon-2025", "entry_id": "liu-et-al-moonlight-scalable-muon-2025", "revision": 1, "title": "Moonlight: update scaling and distributed Muon training", "context": "Large-language-model optimization systems", "track": "Human-designed research method or resource", "band": "B", "modifier": "+", "status": "PROVISIONAL_EDITORIAL", "assessed_on": "2026-10-16", "evidence_cutoff": "2026-10-16", "achievement_state": "SUPPORTED_SCOPED_RESULT", "what_changed": "Added weight decay and adjusted parameter-update scales for large-model Muon, implementing ZeRO-style distributed orthogonalization and releasing Moonlight checkpoints with scaling experiments.", "scope": "Added weight decay and adjusted parameter-update scales for large-model Muon, implementing ZeRO-style distributed orthogonalization and releasing Moonlight checkpoints with scaling experiments. Extends rather than invents Muon/NewtonSchulz/ZeRO, and differs from KimiK2 MuonClip. Approximately twofold compute efficiency is fitted/reported for studied training regimes, not universal wall-clock superiority or tuning-free performance for every architecture.", "evidence_summary": "The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Extends rather than invents Muon/NewtonSchulz/ZeRO, and differs from KimiK2 MuonClip. Approximately twofold compute efficiency is fitted/reported for studied training regimes, not universal wall-clock superiority or tuning-free performance for every architecture.", "references": [ { "title": "Muon is Scalable for LLM Training", "url": "https://arxiv.org/abs/2502.16982v1", "locator": "Original contribution and stated assumptions, evaluation scope and limitations; source-linked assessment, not independent execution or proof certification." } ], "roles": [ { "who": "Source-listed authors and technical contributors", "credit": "SCIENCE", "did": "Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities." } ], "target": { "record_id": "urn:uuid:5fada9ce-dc6c-482d-a63e-2b9921a89d3c", "version": 1, "content_hash": "sha256:87ee892e0bc4f7c9a26c59ecf5bbec4bf1ca6d90fabfb48906fdf61c94b56463" }, "scientific_correctness_certified": false, "independent_expert_calibration": false, "method": "PROPRIETARY", "points": "70000" }, "canonical_agent_submission": false, "scientific_verification_awarded": false, "reader_guide": null, "reading_status": "SOURCE_LINKED_NOTE" }