logbook.json

{
  "schema_version": 1,
  "title": "Reproduction: Generalization and Forgetting in In-Context Continual Learning",
  "emoji": "🧠",
  "space_id": "snaykey/repro-incontext-cl",
  "paper": {
    "openreview_id": "68AMoK2YNk"
  },
  "tags": [
    "icml2026-repro",
    "paper-68AMoK2YNk"
  ],
  "updated_at": "2026-08-02T00:00:00+00:00",
  "root": {
    "slug": "index",
    "title": "Reproduction: Generalization and Forgetting in In-Context Continual Learning",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "claim-1-thm43-decomposition",
        "title": "Theorem 4.3 decomposes the task-t prediction error into an irreducible error term, a variance term scaling as O(M/(t^2(M+1)^2)) that decreases with more in-context examples, and a bias term measuring deviation ‖(1/t)∑_s w_s - w_t‖^2 from task dissimilarity (Theorem 4.3).",
        "file": "pages/claim-1-thm43-decomposition/page.md",
        "children": []
      },
      {
        "slug": "claim-2-thm44-reweighting",
        "title": "Theorem 4.4 shows that forgetting arises from attention reweighting past tasks with negative coefficients c_t and future tasks with positive coefficients d, rather than from actual information loss, comprising intra-task variance and inter-task mean-interaction terms (Theorem 4.4).",
        "file": "pages/claim-2-thm44-reweighting/page.md",
        "children": []
      },
      {
        "slug": "claim-3-nonmonotonic-peak",
        "title": "The theory predicts non-monotonic error curves with context length M, where longer prompts reduce variance but amplify systematic bias from misaligned historical tasks, producing a clear performance peak at intermediate M (Section 4, theoretical analysis).",
        "file": "pages/claim-3-nonmonotonic-peak/page.md",
        "children": []
      },
      {
        "slug": "claim-4-forgetting-floor",
        "title": "Mean-misalignment interference persists regardless of context length, while variance-induced interference decays as O(1/M), implying a persistent asymptotic forgetting floor (Section 4, theoretical analysis).",
        "file": "pages/claim-4-forgetting-floor/page.md",
        "children": []
      },
      {
        "slug": "claim-5-gpt2-validation",
        "title": "Experiments on GPT-2 architectures confirm the predicted non-monotonic error curves and reproduce catastrophic forgetting with a 46% accuracy drop on real LLM tasks (Section 5, experimental validation).",
        "file": "pages/claim-5-gpt2-validation/page.md",
        "children": []
      },
      {
        "slug": "executive-summary",
        "title": "executive-summary",
        "file": "pages/executive-summary/page.md",
        "children": []
      },
      {
        "slug": "conclusion",
        "title": "conclusion",
        "file": "pages/conclusion/page.md",
        "children": []
      }
    ]
  }
}