logbook.json

{
  "schema_version": 1,
  "title": "Reproduction: Provably Learning Attention with Queries",
  "emoji": "🔎",
  "space_id": "snaykey/repro-learning-attention-queries",
  "paper": {
    "openreview_id": "gC04O9t1WJ"
  },
  "tags": [
    "icml2026-repro",
    "paper-gC04O9t1WJ"
  ],
  "updated_at": "2026-08-01T00:00:00+00:00",
  "root": {
    "slug": "index",
    "title": "Reproduction: Provably Learning Attention with Queries",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "claim-1-theorem-4-1-exact-recovery",
        "title": "Single-head softmax attention parameters can be exactly recovered using O(d^2) value queries via a two-phase algorithm that first recovers the output projection vector v with length-1 queries, then recovers the attention matrix W column-by-column with length-2 queries (Theorem 4.1).",
        "file": "pages/claim-1-theorem-4-1-exact-recovery/page.md",
        "children": []
      },
      {
        "slug": "claim-2-theorem-5-2-low-rank",
        "title": "When rank(W) <= r << d, the attention matrix can instead be recovered with only O(rd) queries using rank-one projections and nuclear norm minimization from compressed sensing (Theorem 5.2).",
        "file": "pages/claim-2-theorem-5-2-low-rank/page.md",
        "children": []
      },
      {
        "slug": "claim-3-theorem-6-2-noisy",
        "title": "Under norm bounds and margin conditions on the parameters, epsilon-accurate approximate recovery is achievable with O(d^2) queries tolerating additive noise tau = O(min{mu, epsilon_v/sqrt(d), mu*epsilon_w/(W^2 d)}) (Theorem 6.2).",
        "file": "pages/claim-3-theorem-6-2-noisy/page.md",
        "children": []
      },
      {
        "slug": "claim-4-proposition-7-1-nonidentifiable",
        "title": "Multi-head attention parameters are non-identifiable from queries: distinct parameter sets can produce identical input-output maps, ruling out exact parameter extraction without further structural assumptions (Proposition 7.1).",
        "file": "pages/claim-4-proposition-7-1-nonidentifiable/page.md",
        "children": []
      },
      {
        "slug": "claim-5-sigmoid-reduction",
        "title": "Length-2 queries reduce the softmax attention weight to a sigmoid, alpha = sigma(s1 - s2), enabling inversion via sigma^-1 into linear equations in the unknown parameters (Section preceding Theorem 4.1).",
        "file": "pages/claim-5-sigmoid-reduction/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": []
      }
    ]
  }
}