{
  "arxiv_id": "2605.06216v1",
  "status": "available",
  "source_url": "https://arxiv.org/src/2605.06216v1",
  "retrieved_url": "https://arxiv.org/src/2605.06216v1",
  "retrieved_at_utc": "2026-09-14T05:55:06.831798+00:00",
  "archive_file": "arxiv-source.tar.gz",
  "archive_sha256": "abddce841152dbf8c9c9ad0ddde7f6e3358ae3e51c5cf7e15813099b8c0a5d01",
  "files": [
    {
      "path": "00README.json",
      "bytes": 865,
      "sha256": "2f11bb2b0babe3dcf0d5b1e98ec6e9d585a75a3ef638ba440e5842f52a7cd6d5"
    },
    {
      "path": "apple_preamble.tex",
      "bytes": 2262,
      "sha256": "74d27423628a3dc2e7256c697a49c55b9f34c2d1c42f19a33940af1634ceca40"
    },
    {
      "path": "applemlr.cls",
      "bytes": 6226,
      "sha256": "007ce7c010fc796b2052b29a90ddd55eddbc4982a5c074deef2bdd595e5bf89c"
    },
    {
      "path": "figures/base_contextual_collapse.pdf",
      "bytes": 206522,
      "sha256": "c1930e9fc0c9cb92dbaa0ebc912f6d4bacf6d33b9e1d310bb1f96eeffa6dd01a"
    },
    {
      "path": "figures/contextual_collapse.pdf",
      "bytes": 210877,
      "sha256": "b7bcc7752a90a7cb08507d047e17212f6b918b1805ef8dfd4789581f6cb5066f"
    },
    {
      "path": "figures/embedding_cosine_distance.pdf",
      "bytes": 134340,
      "sha256": "1eeba74bd1012b77928261d362db832ffddc7ddedf4a70174a9e6cc4b708231c"
    },
    {
      "path": "figures/embedding_norm.pdf",
      "bytes": 49426,
      "sha256": "a4e2eaf7132a79ed76737802427a83dd56a9e7e4c6238a80bf6f156f5d86f493"
    },
    {
      "path": "figures/layer_ablation_ppl.pdf",
      "bytes": 30355,
      "sha256": "1a27c6d438f2975689d88db7bc4204c4ac32f57b4e5d18a1a90966b7628f1307"
    },
    {
      "path": "figures/loss_K_plot.pdf",
      "bytes": 27764,
      "sha256": "a1743f51f8333cb92b089fd2e08e1c42b6c526adcb0d5356b7f62b747a418e0e"
    },
    {
      "path": "figures/loss_K_plot_350b.pdf",
      "bytes": 27771,
      "sha256": "5c6cfdacc916f54a89f3722b8c010b67a61b0f2fc127b2a167c8f20b8ee84ce5"
    },
    {
      "path": "figures/main_diagram.pdf",
      "bytes": 97043,
      "sha256": "f05e68fee5923e46162f8f418b508caf907703ecdb093eb56d4daaa9c3a81124"
    },
    {
      "path": "figures/ppl_performance.pdf",
      "bytes": 35593,
      "sha256": "b5919011440a7567ef07b6ce91ebf60839afa53df63f4362ed9afd9f451d0102"
    },
    {
      "path": "figures/rank_reduction_ppl.pdf",
      "bytes": 19041,
      "sha256": "846f5684782511619cd0e881de2fe9ca1ca2b8fb1beb1ebbe7bce6bda58a85db"
    },
    {
      "path": "figures/rare_embedding_analysis_plot.pdf",
      "bytes": 19928,
      "sha256": "96f674195b0b9ef0336fbbfbc5c5481c57a6b5e89d4e2fce48412a6b304c6410"
    },
    {
      "path": "figures/rare_token_loss.pdf",
      "bytes": 35712,
      "sha256": "fab90f771ee51cc1b63a27f636b743aa2fb22dd6fce7a3de2b271296677a332c"
    },
    {
      "path": "figures/rare_token_loss_350b.pdf",
      "bytes": 37458,
      "sha256": "d04e0ad98a024b4979f2b2a7b9669418b4c97356d5ab5b09d87a30fd1a5e999f"
    },
    {
      "path": "figures/router_last_layer_heatmap.pdf",
      "bytes": 135435,
      "sha256": "79962e511b539468214c1930d1192bbededb20e1c72377289a95c80ec2b8a6e9"
    },
    {
      "path": "figures/router_last_layer_null.pdf",
      "bytes": 32929,
      "sha256": "e2a5263ba3e32ee0815ccbf5d1ace4344500990b628c54c598da026d7663bb92"
    },
    {
      "path": "figures/test_loss.jpg",
      "bytes": 508676,
      "sha256": "255b007ec6549c28312c470980e5e1ff931f061e79609b9777b948a1f6b502fd"
    },
    {
      "path": "figures/tide_vs_base_l2norm.pdf",
      "bytes": 46614,
      "sha256": "2557f24466ec069312619cffeb3b1d8df0e2de49694db0de98c5de67e592a1d4"
    },
    {
      "path": "figures/train_loss.jpg",
      "bytes": 207288,
      "sha256": "ef6eefe7b484a5eb6004d20a75a802697a9cc28da01e30cf2b877fca0fde0141"
    },
    {
      "path": "figures/training_loss.pdf",
      "bytes": 2261415,
      "sha256": "a7a4066e0d77639c9213c03d94abd1e61cbfcad06da33db554e121044bbb4fa1"
    },
    {
      "path": "figures/vram_ssd.pdf",
      "bytes": 25081,
      "sha256": "2a20567740358142677acecf1d50a6392329aa476fd91513d7cb8912c0dc587d"
    },
    {
      "path": "main_file.tex",
      "bytes": 93776,
      "sha256": "c95c49a1fe6326444d423a102534783abdca862f5a1f2daa961e2a709b68cea0"
    },
    {
      "path": "math_commands.tex",
      "bytes": 12284,
      "sha256": "90473c4d0542070db244cea73ef962d6cddc5b2a746757e6a40ddf5fdfb90ba9"
    },
    {
      "path": "plainnat.bst",
      "bytes": 26956,
      "sha256": "66447365b962e36bdb18069a1cbd844a8e736ed3201727b69b050c319b2d49f0"
    },
    {
      "path": "references.bib",
      "bytes": 42919,
      "sha256": "32a518963c9e266d9780c3778f8dbf46a23b5a97d9362df9850a982367a4a032"
    },
    {
      "path": "store/apple_logo.pdf",
      "bytes": 7738,
      "sha256": "4174fbdac84c6dd21d27a1fee16269870f035563708323c372f80dfbfb4da37b"
    }
  ],
  "tex_files": [
    "apple_preamble.tex",
    "main_file.tex",
    "math_commands.tex"
  ],
  "entry_candidates": [
    "main_file.tex"
  ],
  "figure_files": [
    "figures/base_contextual_collapse.pdf",
    "figures/contextual_collapse.pdf",
    "figures/embedding_cosine_distance.pdf",
    "figures/embedding_norm.pdf",
    "figures/layer_ablation_ppl.pdf",
    "figures/loss_K_plot.pdf",
    "figures/loss_K_plot_350b.pdf",
    "figures/main_diagram.pdf",
    "figures/ppl_performance.pdf",
    "figures/rank_reduction_ppl.pdf",
    "figures/rare_embedding_analysis_plot.pdf",
    "figures/rare_token_loss.pdf",
    "figures/rare_token_loss_350b.pdf",
    "figures/router_last_layer_heatmap.pdf",
    "figures/router_last_layer_null.pdf",
    "figures/test_loss.jpg",
    "figures/tide_vs_base_l2norm.pdf",
    "figures/train_loss.jpg",
    "figures/training_loss.pdf",
    "figures/vram_ssd.pdf",
    "store/apple_logo.pdf"
  ]
}