logbook.json

{
  "schema_version": 1,
  "title": "Repro: Neural Logistic Bandits",
  "emoji": "chart",
  "space_id": "snaykey/repro-skilltrojan",
  "paper": {
    "openreview_id": "vT5PBMLBvx"
  },
  "tags": [
    "icml2026-repro",
    "paper-vT5PBMLBvx"
  ],
  "updated_at": "2026-07-26T00:00:00+00:00",
  "root": {
    "slug": "index",
    "title": "Repro: Neural Logistic Bandits",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "executive-summary",
        "title": "Executive summary",
        "file": "pages/executive-summary/page.md",
        "children": []
      },
      {
        "slug": "claim-1-neurallog-ucb1-regret-bound",
        "title": "NeuralLog-UCB-1 achieves a regret upper bound of order Õ(d̃√(κT)), where d̃ is the data-adaptive effective dimension rather than the ambient feature dimension d, improving on the d√(κT) bound of Logistic-UCB-1 (Table 1).",
        "file": "pages/claim-1-neurallog-ucb1-regret-bound/page.md",
        "children": []
      },
      {
        "slug": "claim-2-neurallog-ucb2-regret-bound",
        "title": "NeuralLog-UCB-2 achieves a regret upper bound of order Õ(d̃√(T/κ*)), matching the κ* dependence of ada-OFU-ECOLog (whose bound is d√(T/κ*)) while replacing the ambient dimension d with the effective dimension d̃ (Table 1).",
        "file": "pages/claim-2-neurallog-ucb2-regret-bound/page.md",
        "children": []
      },
      {
        "slug": "claim-3-bernstein-martingale-inequality",
        "title": "A novel Bernstein-type self-normalized martingale inequality (Theorem 3.1) yields concentration of order Õ(√d̃) without the extra κ factor required by the corresponding inequality in Verma et al. (2024), which scales as Õ(√(κd̃)) (Theorem 3.1).",
        "file": "pages/claim-3-bernstein-martingale-inequality/page.md",
        "children": []
      },
      {
        "slug": "claim-4-neural-variance-estimates",
        "title": "NeuralLog-UCB-2 uses neural-network-based variance estimates in place of the true reward variance μ̇(xᵀθ*) within the design matrix, enabling the tighter T/κ* dependence while retaining d̃-scaling (Section on NeuralLog-UCB-2 / Condition 5.4).",
        "file": "pages/claim-4-neural-variance-estimates/page.md",
        "children": []
      },
      {
        "slug": "claim-5-synthetic-and-real-experiments",
        "title": "On synthetic nonlinear reward functions h1(x)=0.2(xᵀθ)^4, h2(x)=20cos(xᵀθ), and h3(x)=5xᵀθx, as well as on the MNIST, mushroom, and shuttle real-world datasets, NeuralLog-UCB-2 substantially outperforms baselines including the method of Verma et al. (2024) (Experimental validation section).",
        "file": "pages/claim-5-synthetic-and-real-experiments/page.md",
        "children": []
      },
      {
        "slug": "conclusion",
        "title": "Conclusion",
        "file": "pages/conclusion/page.md",
        "children": []
      }
    ]
  }
}