{
  "arxiv_id": "2609.10715v1",
  "status": "available",
  "source_url": "https://arxiv.org/src/2609.10715v1",
  "retrieved_url": "https://arxiv.org/src/2609.10715v1",
  "retrieved_at_utc": "2026-09-28T09:16:48.540919+00:00",
  "archive_file": "arxiv-source.tar.gz",
  "archive_sha256": "e77ca7af7040e58ab324f8d92b88840fa8ab33f324acebfeb0579ef2b904633c",
  "files": [
    {
      "path": "00README.json",
      "bytes": 213,
      "sha256": "d1f6b1a71c9accb812ec5121d1fff7910c025ca7bb201948bacdb40941bc14e5"
    },
    {
      "path": "appendix.tex",
      "bytes": 46032,
      "sha256": "8d03927926f1274948c1ee85b62dbd49eeb187d57880775f0c54d712e4a9d389"
    },
    {
      "path": "assets/github-icon.png",
      "bytes": 4804,
      "sha256": "c277aa0a90a16bef738e76830e6a0a15df79396052f4dd44afc09ef58dc42437"
    },
    {
      "path": "assets/hf.pdf",
      "bytes": 24570,
      "sha256": "97b71865e5e348d5c61db45f4912a484cef87750983c0b1a8139ddc31d16f24e"
    },
    {
      "path": "assets/huggingface-icon.png",
      "bytes": 9925,
      "sha256": "5ee3de50954b62529f2e1e3ca885c825f7247b4b24b8a5056ac0e2bf5eab6c0c"
    },
    {
      "path": "assets/logo.png",
      "bytes": 1162825,
      "sha256": "cbc20e87896a9e25abe2c420dd946248ac6bd0d4a1418e1dbc98cee23e7a4887"
    },
    {
      "path": "figure/NCP-ArchPreview_main_V2.pdf",
      "bytes": 882643,
      "sha256": "fa4325c586044a3d18b2748ae9c1206c5a75eff286a8adab34ac58db08a1bb96"
    },
    {
      "path": "figure/ablation.pdf",
      "bytes": 652799,
      "sha256": "ff915e6ba2afbfec7efa8a7cd8b146c4c9a1024a57f160d6257b8b49dff6b46a"
    },
    {
      "path": "figure/conceptlm_3e18_curve_pretty_from_table.pdf",
      "bytes": 18047,
      "sha256": "d5c823c0de49f4c7c83425e568ca5e9bf6dda62f098006076bed8b9e1187362e"
    },
    {
      "path": "figure/d_proxy_downstream_fit_100m_concept_v123.pdf",
      "bytes": 39096,
      "sha256": "b14250d75ac321a1fe36efa565c97d3283ccd48c2abe15bd378162c29df30c33"
    },
    {
      "path": "figure/max_logits_and_qk_norm.pdf",
      "bytes": 14092,
      "sha256": "5cc7f2989c7c70ecca734427334e975a5d3579342a9c4be3163e0af6c3d9ce43"
    },
    {
      "path": "figure/max_logits_grad_norm_qk_norm_triptych.pdf",
      "bytes": 25894,
      "sha256": "0b868ef36c2c1f8408923212502b15ef79e580c13d05773a4d78ff27569882a9"
    },
    {
      "path": "figure/ncp_archpreview_mtp_loss_updates_ema200.pdf",
      "bytes": 294107,
      "sha256": "7f7509d2d83b646166030977441e4f876e2cad6c8c02f5823e497a5d4630328c"
    },
    {
      "path": "figure/scaling-law-nd.pdf",
      "bytes": 10065,
      "sha256": "90ce949b15638f079ac02a42d7d7d115b1b6a39a4ccda7fa404b820f29e392f6"
    },
    {
      "path": "figure/stage1_loss.pdf",
      "bytes": 275024,
      "sha256": "5d6ca98dfc20363147db89e386005d4b66550518dbf77df68a9172dd5e59c726"
    },
    {
      "path": "figure/stage2_loss.pdf",
      "bytes": 85512,
      "sha256": "7f887278e6187b8fec7c4d73e20637d039aedefdcc5b6b0d1a727a8c773ce29c"
    },
    {
      "path": "figure/training_efficiency_gpu_memory.pdf",
      "bytes": 23251,
      "sha256": "5bc2070ef16a2bd1090db98d606bf5086a6ead97cb708c84141b97335c22d010"
    },
    {
      "path": "lumia.cls",
      "bytes": 12963,
      "sha256": "0b4d721041a8f239c058fde50858a3100a5805ced07bf7b9a05e8808ad69d038"
    },
    {
      "path": "main.tex",
      "bytes": 22682,
      "sha256": "4582989aa57118fea5c3eb6838c2789403aabb5167f0b214536f6fcfc9e253fe"
    },
    {
      "path": "notations.tex",
      "bytes": 3849,
      "sha256": "49ea0539d9ea18230c750d2582c54cc3ab13f04897fccf8d89da9b3a2b334111"
    },
    {
      "path": "references.bib",
      "bytes": 46890,
      "sha256": "073068929936438d794ab341baea5ceb2c2f40807940b05ad20408d44427ffc9"
    },
    {
      "path": "sections/01_introduction.tex",
      "bytes": 7639,
      "sha256": "2854005dfe4acdc8257a2eaa5c0a61de8094e4dc281946e8a38c9f01428086b5"
    },
    {
      "path": "sections/02_architecture.tex",
      "bytes": 14349,
      "sha256": "565e4ef717ea326868890402f54e32a5f5957270e2c480391380405fadf03372"
    },
    {
      "path": "sections/03_training.tex",
      "bytes": 6316,
      "sha256": "8d4359dfb7be50163c2ac5291043aa54ee762f353e39241c5b8c29c5758cccd6"
    },
    {
      "path": "sections/04_evaluation.tex",
      "bytes": 30989,
      "sha256": "40655bd7d756ec7e8a7fdb9400e617949cc462d34621dbb2759185164bff6750"
    },
    {
      "path": "sections/05_related_work.tex",
      "bytes": 4316,
      "sha256": "6185bc96255a3078896fc8ec16657e1eb743b51393b42afe368c63c20dbfc56a"
    },
    {
      "path": "sections/07_limitations.tex",
      "bytes": 1893,
      "sha256": "ce58069606ed9b2c56177bdd63dc471f662dca8ae4628b76e5e7457b6f31c685"
    },
    {
      "path": "sections/08_conclusion.tex",
      "bytes": 1531,
      "sha256": "098f417703f0e513c28380054e0e2471a788707e2bf4fa2962732a079fcf5e45"
    },
    {
      "path": "table/scaling-experiments.tex",
      "bytes": 2991,
      "sha256": "1da0c1cee5ffc6be45d3f2dd315ace551c7e4c7a3d95c6818b082b5bb010434e"
    },
    {
      "path": "table/table-hlm_mtp_results.tex",
      "bytes": 1045,
      "sha256": "4aa6aa69e9d5fa800ce17acd5c3e631661e95a772eb97463f1ed0b9ed5df778c"
    },
    {
      "path": "table/table-main-results.tex",
      "bytes": 6623,
      "sha256": "a99ac526a00153d161a74ce4ebed08e9fe6fd42558b0143585e4e7c615f7cc9f"
    },
    {
      "path": "table/table-vq-code-results.tex",
      "bytes": 6025,
      "sha256": "440a80a4d0ca4dae71429cace1cb76ddf0ca89f8e3077999dd3900c144805b58"
    },
    {
      "path": "table/table-vq-knowledge-results.tex",
      "bytes": 4242,
      "sha256": "d4ad6942252b284f5192511fd088a564c6b4aca398f0ef09c8844459fe6e6b24"
    },
    {
      "path": "table/table-vq-math-results.tex",
      "bytes": 5232,
      "sha256": "8ca84312c9ad5df733eb569a13360947ba444e3604b406e1826a00e758ec1966"
    },
    {
      "path": "table/table-vq-parameter-cost.tex",
      "bytes": 542,
      "sha256": "68bf7ec26a0e00419cfbb6b056aa8612a0dab818d168030d3d8fdb0b71c77124"
    }
  ],
  "tex_files": [
    "appendix.tex",
    "main.tex",
    "notations.tex",
    "sections/01_introduction.tex",
    "sections/02_architecture.tex",
    "sections/03_training.tex",
    "sections/04_evaluation.tex",
    "sections/05_related_work.tex",
    "sections/07_limitations.tex",
    "sections/08_conclusion.tex",
    "table/scaling-experiments.tex",
    "table/table-hlm_mtp_results.tex",
    "table/table-main-results.tex",
    "table/table-vq-code-results.tex",
    "table/table-vq-knowledge-results.tex",
    "table/table-vq-math-results.tex",
    "table/table-vq-parameter-cost.tex"
  ],
  "entry_candidates": [
    "main.tex"
  ],
  "figure_files": [
    "assets/github-icon.png",
    "assets/hf.pdf",
    "assets/huggingface-icon.png",
    "assets/logo.png",
    "figure/NCP-ArchPreview_main_V2.pdf",
    "figure/ablation.pdf",
    "figure/conceptlm_3e18_curve_pretty_from_table.pdf",
    "figure/d_proxy_downstream_fit_100m_concept_v123.pdf",
    "figure/max_logits_and_qk_norm.pdf",
    "figure/max_logits_grad_norm_qk_norm_triptych.pdf",
    "figure/ncp_archpreview_mtp_loss_updates_ema200.pdf",
    "figure/scaling-law-nd.pdf",
    "figure/stage1_loss.pdf",
    "figure/stage2_loss.pdf",
    "figure/training_efficiency_gpu_memory.pdf"
  ]
}