{
  "schema_version": 1,
  "title": "Repro: NorMuon: Making Muon more efficient and scalable",
  "emoji": "🎯",
  "space_id": "alphaXiv/m1IRWFAMsa",
  "paper": {
    "openreview_id": "m1IRWFAMsa",
    "claims": [
      {
        "status": "reproduced"
      },
      {
        "status": "reproduced"
      },
      {
        "status": "unverified"
      },
      {
        "status": "unverified"
      },
      {
        "status": "reproduced"
      },
      {
        "status": "in-progress"
      }
    ]
  },
  "tags": [
    "icml2026-repro",
    "paper-m1IRWFAMsa"
  ],
  "updated_at": "2026-07-09T11:32:53+00:00",
  "root": {
    "slug": "index",
    "title": "Repro: NorMuon: Making Muon more efficient and scalable",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "claim-1-normuon-algorithm-muon-orthogonalization-neuron-wise-2nd-moment-normalization-alg-1",
        "title": "Claim 1: NorMuon algorithm = Muon orthogonalization + neuron-wise 2nd-moment normalization (Alg. 1)",
        "file": "pages/claim-1-normuon-algorithm-muon-orthogonalization-neuron-wise-2nd-moment-normalization-alg-1/page.md",
        "children": []
      },
      {
        "slug": "claim-2-muon-leaves-high-per-neuron-update-norm-variance-normuon-normalizes-it-fig-1",
        "title": "Claim 2: Muon leaves high per-neuron update-norm variance; NorMuon normalizes it (Fig. 1)",
        "file": "pages/claim-2-muon-leaves-high-per-neuron-update-norm-variance-normuon-normalizes-it-fig-1/page.md",
        "children": []
      },
      {
        "slug": "claim-3-21-74-step-efficiency-gain-over-adam-11-31-over-muon-at-1-1b-table-1",
        "title": "Claim 3: 21.74% step-efficiency gain over Adam, 11.31% over Muon at 1.1B (Table 1)",
        "file": "pages/claim-3-21-74-step-efficiency-gain-over-adam-11-31-over-muon-at-1-1b-table-1/page.md",
        "children": []
      },
      {
        "slug": "claim-4-normuon-improves-validation-loss-trajectories-at-5-4b-fig-2",
        "title": "Claim 4: NorMuon improves validation-loss trajectories at 5.4B (Fig. 2)",
        "file": "pages/claim-4-normuon-improves-validation-loss-trajectories-at-5-4b-fig-2/page.md",
        "children": []
      },
      {
        "slug": "claim-5-optimizer-memory-near-muon-2-9-step-time-overhead-vs-adamw-table-2",
        "title": "Claim 5: Optimizer memory near Muon, ~2.9% step-time overhead vs AdamW (Table 2)",
        "file": "pages/claim-5-optimizer-memory-near-muon-2-9-step-time-overhead-vs-adamw-table-2/page.md",
        "children": []
      },
      {
        "slug": "claim-6-normuon-outperforms-muon-in-modded-nanogpt-at-124m-and-350m-fig-5",
        "title": "Claim 6: NorMuon outperforms Muon in Modded-NanoGPT at 124M and 350M (Fig. 5)",
        "file": "pages/claim-6-normuon-outperforms-muon-in-modded-nanogpt-at-124m-and-350m-fig-5/page.md",
        "children": []
      }
    ]
  },
  "agent_view_tokens": 3925,
  "revision": "1783596773111378000"
}