{
  "arxiv_id": "2602.08984v1",
  "status": "available",
  "source_url": "https://arxiv.org/src/2602.08984v1",
  "retrieved_url": "https://arxiv.org/src/2602.08984v1",
  "retrieved_at_utc": "2026-09-28T08:33:06.922885+00:00",
  "archive_file": "arxiv-source.tar.gz",
  "archive_sha256": "ad1cf8d9db8bee4450e7f6e8e1593c51d75715947a1963d555c2c2adc31423d8",
  "files": [
    {
      "path": "00README.json",
      "bytes": 213,
      "sha256": "d1f6b1a71c9accb812ec5121d1fff7910c025ca7bb201948bacdb40941bc14e5"
    },
    {
      "path": "assets/hf.pdf",
      "bytes": 24570,
      "sha256": "97b71865e5e348d5c61db45f4912a484cef87750983c0b1a8139ddc31d16f24e"
    },
    {
      "path": "assets/logo.png",
      "bytes": 1162825,
      "sha256": "cbc20e87896a9e25abe2c420dd946248ac6bd0d4a1418e1dbc98cee23e7a4887"
    },
    {
      "path": "figs/Long_range_dependency.pdf",
      "bytes": 16929,
      "sha256": "b7ac2302583b1e80eb8790111c4cf3c638be7c806db22e7004274e9a0ae30b49"
    },
    {
      "path": "figs/VQ_frequency.pdf",
      "bytes": 1279106,
      "sha256": "fe3aff8eee81202d1d190463e969d910d79b6fdccbc1a53dd47a68dca6c087c1"
    },
    {
      "path": "figs/flop_scaling.pdf",
      "bytes": 16745,
      "sha256": "04fa50cd668a4c2eaf1e8943b93c58e45525c3615a8c1a5e7981f72b0698c859"
    },
    {
      "path": "figs/gradient_flow_V4.pdf",
      "bytes": 96071,
      "sha256": "9bb2154301bcbec2ce45eedcea868eaceb78307a0e128c56b0e4429906d4ce5a"
    },
    {
      "path": "figs/main_arc.pdf",
      "bytes": 517972,
      "sha256": "c43757ff1a345fd1a30d578315f639ee9445e641689e2066c8108fe5914ff225"
    },
    {
      "path": "figs/param_scaling.pdf",
      "bytes": 16984,
      "sha256": "570489828f42e5a18171642459d5dc875aa1ada65d32aae6f779b561390f5a5c"
    },
    {
      "path": "figs/token_scaling.pdf",
      "bytes": 16790,
      "sha256": "49761016e5ab4a00e3cdf4dccd1b6d0999a6ddc414626ceb0e598ddf6a38869a"
    },
    {
      "path": "files/1-intro.tex",
      "bytes": 5488,
      "sha256": "ef06383c711a37a9bfe8dc0c33ca4614c5e7a95b39d50ece3bbb3e5936bb8275"
    },
    {
      "path": "files/2-related_work.tex",
      "bytes": 2779,
      "sha256": "3ba8c87bb4f3bcebdaf212e80d0f1a9d599c498110faf4e5703552e78e8bb169"
    },
    {
      "path": "files/3-method.tex",
      "bytes": 8980,
      "sha256": "bb6621e0e25448a6be0517677d1957c78cef83eda45b1e6f1b8c358c791295f1"
    },
    {
      "path": "files/4-exp.tex",
      "bytes": 8205,
      "sha256": "b036a0489a476e94c5f4abe7861bbf26372696a58bce59614e9c0668831d3d42"
    },
    {
      "path": "files/5-analysis.tex",
      "bytes": 6435,
      "sha256": "1cf9fcccaf916a8ad67c4e02fff0e4d71bf9257462e746a0fde81bc660a46c19"
    },
    {
      "path": "files/6-conclusion.tex",
      "bytes": 1510,
      "sha256": "e1dce8e763f8a96ef391776207d6956bff57e2af5ce4a26368ec76e4ce7bdcb3"
    },
    {
      "path": "files/7-appendix.tex",
      "bytes": 11168,
      "sha256": "51d5e9146d0bd94876e57273b185c4cd04517d71774533539942ec165f659205"
    },
    {
      "path": "files/tables/App_table_GPT2_withL1.tex",
      "bytes": 8047,
      "sha256": "87e397c76468c53eac2794fe8692bf9ce04ec0d1999c54952220611cb634f7cf"
    },
    {
      "path": "files/tables/Chunk_size_abl.tex",
      "bytes": 2109,
      "sha256": "ad42d80b44afa247eeb9bea720768aded5a3cef18c0f2f60de78ccf273279b9c"
    },
    {
      "path": "files/tables/GPT_Phthia_ppl_tab.tex",
      "bytes": 3263,
      "sha256": "327c28eabd33de4a5c15cf7959ef2cdd676357417470a3a848f0fe33ed4cbad1"
    },
    {
      "path": "files/tables/Insert_layer.tex",
      "bytes": 2919,
      "sha256": "8b076d777dcbc33be8efcd700c7651ec2cbd722a93d57882aa530839dd77cc62"
    },
    {
      "path": "files/tables/Llama3_training_infer_time.tex",
      "bytes": 1073,
      "sha256": "48ccc3c96a939f27fbe905fd25dd49a9afcf74fd6278a9bcca2bbaf09dd4dd5a"
    },
    {
      "path": "files/tables/Llama_res_tab.tex",
      "bytes": 1240,
      "sha256": "0ce10980bba2d23f5403f62f9af47cbe5dc84ec857ea639a55d61c6a8cf05177"
    },
    {
      "path": "files/tables/NTP_vs_NCP.tex",
      "bytes": 1042,
      "sha256": "abd7cddd2d1931e5d26a2c17ebd5caecdcfce633e0c9ef95b15dd7d637b28cd0"
    },
    {
      "path": "files/tables/Pythia_ppl_tab.tex",
      "bytes": 4623,
      "sha256": "41c0a7f076469943e593f8b16365f2f44651f019eca6828fdfb2124910c00452"
    },
    {
      "path": "files/tables/VQ_S_abl.tex",
      "bytes": 3107,
      "sha256": "806bc54c355c833ca5da4c40cba7c5110bd8cbb6521f6f2925788612c7525e13"
    },
    {
      "path": "files/tables/VQ_size.tex",
      "bytes": 2578,
      "sha256": "b145c218639751bc73e47dd6628ab59f16fb9f7d38a26c4c0b578e8ee40180f5"
    },
    {
      "path": "files/tables/loss_ablation_detailed.tex",
      "bytes": 3020,
      "sha256": "8d8b54d77c5a669bc9b96ee4a97826d764d9adeb56b29552504045dc6483d767"
    },
    {
      "path": "files/tables/scaling_res_detailed.tex",
      "bytes": 3105,
      "sha256": "35e26b4becf177f48d82cb51a183e599b76696a1b3e5f08e90ee5ac7fa06465f"
    },
    {
      "path": "lumia.cls",
      "bytes": 12963,
      "sha256": "0b4d721041a8f239c058fde50858a3100a5805ced07bf7b9a05e8808ad69d038"
    },
    {
      "path": "main.bib",
      "bytes": 24329,
      "sha256": "d984cd17d33c888fa6a58f3d28e12f3c476057babac6d5737693ded8671047a0"
    },
    {
      "path": "main.tex",
      "bytes": 20268,
      "sha256": "543c933a19e5866236f9f06e1859bb4709ec4c5682302098f0da59b97bce600b"
    },
    {
      "path": "notations.tex",
      "bytes": 3849,
      "sha256": "49ea0539d9ea18230c750d2582c54cc3ab13f04897fccf8d89da9b3a2b334111"
    }
  ],
  "tex_files": [
    "files/1-intro.tex",
    "files/2-related_work.tex",
    "files/3-method.tex",
    "files/4-exp.tex",
    "files/5-analysis.tex",
    "files/6-conclusion.tex",
    "files/7-appendix.tex",
    "files/tables/App_table_GPT2_withL1.tex",
    "files/tables/Chunk_size_abl.tex",
    "files/tables/GPT_Phthia_ppl_tab.tex",
    "files/tables/Insert_layer.tex",
    "files/tables/Llama3_training_infer_time.tex",
    "files/tables/Llama_res_tab.tex",
    "files/tables/NTP_vs_NCP.tex",
    "files/tables/Pythia_ppl_tab.tex",
    "files/tables/VQ_S_abl.tex",
    "files/tables/VQ_size.tex",
    "files/tables/loss_ablation_detailed.tex",
    "files/tables/scaling_res_detailed.tex",
    "main.tex",
    "notations.tex"
  ],
  "entry_candidates": [
    "main.tex"
  ],
  "figure_files": [
    "assets/hf.pdf",
    "assets/logo.png",
    "figs/Long_range_dependency.pdf",
    "figs/VQ_frequency.pdf",
    "figs/flop_scaling.pdf",
    "figs/gradient_flow_V4.pdf",
    "figs/main_arc.pdf",
    "figs/param_scaling.pdf",
    "figs/token_scaling.pdf"
  ]
}