logbook.json

{
  "schema_version": 1,
  "title": "Reproduction: Anytime Safe PAC Efficient Reasoning",
  "emoji": "🎰",
  "space_id": "snaykey/repro-pine-conformal-pruning",
  "paper": {
    "openreview_id": "iMuMdWc7uC"
  },
  "tags": [
    "icml2026-repro",
    "paper-iMuMdWc7uC"
  ],
  "updated_at": "2026-07-31T00:00:00+00:00",
  "root": {
    "slug": "index",
    "title": "Reproduction: Anytime Safe PAC Efficient Reasoning",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "claim-1-algorithm1-betting-game",
        "title": "B-PAC reasoning adaptively routes queries between a thinking and non-thinking model by treating threshold selection as a betting game, using inverse-propensity-scoring estimators to build supermartingales that accumulate statistical evidence of safety (Section 3, Algorithm 1).",
        "file": "pages/claim-1-algorithm1-betting-game/page.md",
        "children": []
      },
      {
        "slug": "claim-2-theorem-4-2-anytime-pac",
        "title": "Theorem 4.2 establishes anytime PAC efficiency: the probability that the empirical risk R_t(u_t) stays below epsilon for all time steps t is at least 1-alpha, holding for both i.i.d. and non-stationary data (Theorem 4.2, Sections 4.1-4.2, Section 5).",
        "file": "pages/claim-2-theorem-4-2-anytime-pac/page.md",
        "children": []
      },
      {
        "slug": "claim-3-theorem-4-3-logt-regret",
        "title": "Theorem 4.3 shows the adaptive betting strategy achieves O(log T) regret, giving rapid convergence to the optimal efficient routing threshold (Theorem 4.3).",
        "file": "pages/claim-3-theorem-4-3-logt-regret/page.md",
        "children": []
      },
      {
        "slug": "claim-4-magpie-efficiency",
        "title": "On the Magpie dataset with epsilon=0.08, B-PAC routes only 18.99% of queries to the thinking model while achieving 41.37% token savings relative to always using the thinking model (Section 6.1, Table 1).",
        "file": "pages/claim-4-magpie-efficiency/page.md",
        "children": []
      },
      {
        "slug": "claim-5-mmlupro-efficiency",
        "title": "On MMLU-Pro with epsilon=0.08, B-PAC achieves a 47.04% expert (thinking-model) call percentage and 78.07% token percentage, corresponding to up to an 81.01% reduction in thinking-model usage reported in the abstract (Section 6.1, Table 1).",
        "file": "pages/claim-5-mmlupro-efficiency/page.md",
        "children": []
      },
      {
        "slug": "claim-6-setup-baselines",
        "title": "Experiments use Qwen3-4B-Thinking-2507 as the thinking model and Qwen3-4B-Instruct-2507 as the non-thinking model, evaluated on MATH, MMLU-Pro, BIG-Bench Hard, and Magpie, compared against online baselines IPS+Hoeffding and O-Naive (Section 6, Figure 3, Table 1).",
        "file": "pages/claim-6-setup-baselines/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": []
      }
    ]
  }
}