{
  "schema_version": 1,
  "title": "Repro - Do We Need Adam? Surprisingly Strong and Sparse Reinforcement Learning with SGD in LLMs",
  "emoji": "\ud83d\udd2c",
  "space_id": "bertfil/z31fdV4WRu",
  "paper": {
    "arxiv_id": "2602.07729",
    "openreview_id": "z31fdV4WRu",
    "title": "Do We Need Adam? Surprisingly Strong and Sparse Reinforcement Learning with SGD in LLMs",
    "url": "https://arxiv.org/abs/2602.07729"
  },
  "tags": [
    "icml2026-repro",
    "paper-z31fdV4WRu"
  ],
  "updated_at": "2026-07-20T03:52:33+00:00",
  "root": {
    "slug": "index",
    "title": "Repro - Do We Need Adam? Surprisingly Strong and Sparse Reinforcement Learning with SGD in LLMs",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "00-scored-evidence-summary",
        "title": "00 - Scored evidence summary",
        "file": "pages/00-scored-evidence-summary/page.md",
        "children": []
      },
      {
        "slug": "claim-1-on-qwen3-8b-math-reasoning-with-3k-response-length",
        "title": "Claim 1 - On Qwen3-8B math reasoning with 3K response length, SGD atta\u2026",
        "file": "pages/claim-1-on-qwen3-8b-math-reasoning-with-3k-response-length/page.md",
        "children": []
      },
      {
        "slug": "claim-2-on-qwen3-1-7b-at-4k-context-sgd-improves-humaneval",
        "title": "Claim 2 - On Qwen3-1.7B at 4K context, SGD improves HumanEval@1 coding\u2026",
        "file": "pages/claim-2-on-qwen3-1-7b-at-4k-context-sgd-improves-humaneval/page.md",
        "children": []
      },
      {
        "slug": "claim-3-sgd-updates-reach-99-99-sparsity-updating-about-0",
        "title": "Claim 3 - SGD updates reach 99.99% sparsity (updating about 0.01% of p\u2026",
        "file": "pages/claim-3-sgd-updates-reach-99-99-sparsity-updating-about-0/page.md",
        "children": []
      },
      {
        "slug": "claim-4-sgd-updates-also-exhibit-substantially-lower-effec",
        "title": "Claim 4 - SGD updates also exhibit substantially lower effective rank \u2026",
        "file": "pages/claim-4-sgd-updates-also-exhibit-substantially-lower-effec/page.md",
        "children": []
      },
      {
        "slug": "claim-5-ablations-decomposing-adamw-into-its-momentum-and",
        "title": "Claim 5 - Ablations decomposing AdamW into its momentum and RMSProp (a\u2026",
        "file": "pages/claim-5-ablations-decomposing-adamw-into-its-momentum-and/page.md",
        "children": []
      },
      {
        "slug": "claim-6-on-extended-ppo-training-rlve-500-steps-on-qwen3-8",
        "title": "Claim 6 - On extended PPO training (RLVE, 500 steps) on Qwen3-8B, SGD \u2026",
        "file": "pages/claim-6-on-extended-ppo-training-rlve-500-steps-on-qwen3-8/page.md",
        "children": []
      },
      {
        "slug": "reproduction-protocol-and-provenance",
        "title": "Reproduction protocol and provenance",
        "file": "pages/reproduction-protocol-and-provenance/page.md",
        "children": []
      },
      {
        "slug": "conclusion",
        "title": "Conclusion",
        "file": "pages/conclusion/page.md",
        "children": []
      }
    ]
  },
  "agent_view_tokens": 9836,
  "revision": "1784519553165232100"
}