{
  "schema_version": 1,
  "title": "Reproduction: Predictive Variational Inference (PVI)",
  "emoji": "🎯",
  "space_id": "snaykey/repro-predictive-vi",
  "paper": {
    "openreview_id": "dVIts0fNb4"
  },
  "tags": [
    "icml2026-repro",
    "paper-dVIts0fNb4"
  ],
  "updated_at": "2026-08-02T00:00:00+00:00",
  "root": {
    "slug": "index",
    "title": "Reproduction: Predictive Variational Inference (PVI)",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "claim-1-framework",
        "title": "Predictive variational inference (PVI) replaces the KL-divergence objective of standard VI with optimization of the posterior predictive distribution against the true data-generating process, using proper scoring rules (Section: Main Framework).",
        "file": "pages/claim-1-framework/page.md",
        "children": []
      },
      {
        "slug": "claim-2-prop1-consistency",
        "title": "Proposition 1 establishes that PVI solutions converge to the predictively optimal parameter as sample size grows, even under model misspecification (Proposition 1).",
        "file": "pages/claim-2-prop1-consistency/page.md",
        "children": []
      },
      {
        "slug": "claim-3-corollaries",
        "title": "Corollary 1 shows PVI recovers the true parameter when the model is correctly specified, and Corollary 2 shows it recovers the population parameter distribution when misspecified with population-varying parameters (Corollary 1, Corollary 2).",
        "file": "pages/claim-3-corollaries/page.md",
        "children": []
      },
      {
        "slug": "claim-4-scoring-rules",
        "title": "PVI is implemented with three proper scoring rules -- logarithmic, quadratic, and continuous ranked probability score (CRPS) -- with the CRPS variant enabling likelihood-free inference for intractable simulators (Section on scoring rules).",
        "file": "pages/claim-4-scoring-rules/page.md",
        "children": []
      },
      {
        "slug": "claim-5-cryoem-heterogeneity",
        "title": "In a cryo-EM protein conformation inference experiment, PVI recovers the true population distribution of parameters while standard Bayesian VI collapses to a point estimate (Section: CryoEM Protein Inference experiment).",
        "file": "pages/claim-5-cryoem-heterogeneity/page.md",
        "children": []
      },
      {
        "slug": "claim-6-posteriordb",
        "title": "Across 7 models from PosteriorDB, PVI improves held-out predictive performance compared to standard variational inference (Section: Benchmark PosteriorDB experiment).",
        "file": "pages/claim-6-posteriordb/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": []
      }
    ]
  }
}