{
  "schema_version": 1,
  "title": "Repro: Stable Deep Reinforcement Learning via Isotropic Gaussian Representations",
  "emoji": "target",
  "space_id": "snaykey/repro-plugmem",
  "paper": {
    "openreview_id": "gc7Gg18ejz"
  },
  "tags": [
    "icml2026-repro",
    "paper-gc7Gg18ejz"
  ],
  "updated_at": "2026-07-26T00:00:00+00:00",
  "root": {
    "slug": "index",
    "title": "Repro: Stable Deep Reinforcement Learning via Isotropic Gaussian Representations",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "executive-summary",
        "title": "Executive summary",
        "file": "pages/executive-summary/page.md",
        "children": []
      },
      {
        "slug": "claim-1-tracking",
        "title": "The paper proves tracking-error dynamics for a linear critic under non-stationary targets and connects stability to embedding covariance geometry (Theorem 3.1).",
        "file": "pages/claim-1-tracking/page.md",
        "children": []
      },
      {
        "slug": "claim-2-isotropy",
        "title": "The analysis argues that isotropic representations equalize contraction directions and reduce drift amplification under a fixed variance budget (Section 3.1).",
        "file": "pages/claim-2-isotropy/page.md",
        "children": []
      },
      {
        "slug": "claim-3-gaussian-moments",
        "title": "The paper motivates Gaussian representations because Gaussian higher-order moments are determined by covariance, limiting uncontrolled tail behavior (Section 3.1).",
        "file": "pages/claim-3-gaussian-moments/page.md",
        "children": []
      },
      {
        "slug": "claim-4-sigreg",
        "title": "Sketched Isotropic Gaussian Regularization shapes learned representations toward an isotropic Gaussian distribution during RL training (Section 4).",
        "file": "pages/claim-4-sigreg/page.md",
        "children": []
      },
      {
        "slug": "claim-5-cifar",
        "title": "On non-stationary CIFAR-10, adding SIGReg improves train accuracy AUC, feature rank, and dormant-neuron metrics across Adam, RAdam, and Kron optimizers (Table 1).",
        "file": "pages/claim-5-cifar/page.md",
        "children": []
      },
      {
        "slug": "claim-6-rl-domains",
        "title": "The empirical study reports that SIGReg reduces representation collapse, neuron dormancy, and training instability across evaluated RL domains (Section 5).",
        "file": "pages/claim-6-rl-domains/page.md",
        "children": []
      },
      {
        "slug": "conclusion",
        "title": "Conclusion",
        "file": "pages/conclusion/page.md",
        "children": []
      }
    ]
  }
}