{
  "schema_version": 1,
  "title": "Repro - How much can language models memorize?",
  "emoji": "🧠",
  "space_id": "abidlabs/memorize",
  "paper": {
    "openreview_id": "bA6BgSbaUi",
    "arxiv_id": "2505.24832",
    "icml_id": "27372",
    "title": "How much can language models memorize?",
    "claims": [
      {
        "text": "GPT-style transformers trained on uniform random data show an empirical memorization-capacity plateau of about 3.6 bits per parameter (Figure 1)",
        "status": "in-progress"
      },
      {
        "text": "Capacity estimates across model widths and depths support a roughly linear bits-per-parameter scaling law, with bfloat16 to float32 increasing capacity only modestly (Table 1)",
        "status": "in-progress"
      },
      {
        "text": "On text data, unintended memorization rises with model size but decreases once models begin generalizing relative to an oracle reference model (Figure 2)",
        "status": "in-progress"
      },
      {
        "text": "Double descent begins when dataset information content exceeds estimated model capacity in both synthetic bitstrings and text experiments (Figures 3 and 4)",
        "status": "in-progress"
      },
      {
        "text": "The paper derives and evaluates scaling-law predictions for membership inference as a function of model capacity and dataset size (Figure 7)",
        "status": "in-progress"
      }
    ]
  },
  "tags": [
    "icml2026-repro",
    "paper-bA6BgSbaUi"
  ],
  "updated_at": "2026-07-11T00:04:34+00:00",
  "root": {
    "slug": "index",
    "title": "Repro - How much can language models memorize?",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "claim-1-3-6-bits-param-plateau-synthetic",
        "title": "Claim 1: 3.6 bits/param plateau (synthetic)",
        "file": "pages/claim-1-3-6-bits-param-plateau-synthetic/page.md",
        "children": []
      },
      {
        "slug": "claim-2-width-depth-scaling-precision",
        "title": "Claim 2: width/depth scaling + precision",
        "file": "pages/claim-2-width-depth-scaling-precision/page.md",
        "children": []
      },
      {
        "slug": "claim-3-text-memorization-vs-generalization",
        "title": "Claim 3: text memorization vs generalization",
        "file": "pages/claim-3-text-memorization-vs-generalization/page.md",
        "children": []
      },
      {
        "slug": "claim-4-double-descent-at-capacity",
        "title": "Claim 4: double descent at capacity",
        "file": "pages/claim-4-double-descent-at-capacity/page.md",
        "children": []
      },
      {
        "slug": "claim-5-membership-inference-scaling",
        "title": "Claim 5: membership inference scaling",
        "file": "pages/claim-5-membership-inference-scaling/page.md",
        "children": []
      },
      {
        "slug": "memorize-synthetic",
        "title": "memorize-synthetic",
        "file": "pages/memorize-synthetic/page.md",
        "children": []
      },
      {
        "slug": "memorize-claim4",
        "title": "memorize-claim4",
        "file": "pages/memorize-claim4/page.md",
        "children": []
      },
      {
        "slug": "memorize-text",
        "title": "memorize-text",
        "file": "pages/memorize-text/page.md",
        "children": []
      }
    ]
  },
  "agent_view_tokens": 6448,
  "revision": "1783728274322498000"
}