{
  "schema_version": 2,
  "title": "Reproduction: Rethinking the Design Space of Reinforcement Learning for Diffusion Models: On the Importance of Likelihood Estimation Beyond Loss Design",
  "emoji": "🎯",
  "space_id": "Firemedic15/repro-rl-design-space-diffusion-models-likelihood-estimation",
  "paper": null,
  "tags": [
    "icml2026-repro",
    "paper-3mJ9eTU319"
  ],
  "updated_at": "2026-07-22T15:53:20+00:00",
  "root": {
    "slug": "index",
    "title": "Reproduction: Rethinking the Design Space of Reinforcement Learning for Diffusion Models: On the Importance of Likelihood Estimation Beyond Loss Design",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "executive-summary",
        "title": "Executive summary",
        "file": "pages/executive-summary/page.md",
        "children": []
      },
      {
        "slug": "claim-1-design-space-deconstruction-elbo-likelihood-estimation-is-the-dominant-factor",
        "title": "Claim 1: Design space deconstruction — ELBO likelihood estimation is the dominant factor",
        "file": "pages/claim-1-design-space-deconstruction-elbo-likelihood-estimation-is-the-dominant-factor/page.md",
        "children": []
      },
      {
        "slug": "claim-2-geneval-0-24-0-95-in-90-gpu-hours-4-6x-over-flowgrpo-2x-over-diffusionnft",
        "title": "Claim 2: GenEval 0.24→0.95 in 90 GPU hours, 4.6x over FlowGRPO, 2x over DiffusionNFT",
        "file": "pages/claim-2-geneval-0-24-0-95-in-90-gpu-hours-4-6x-over-flowgrpo-2x-over-diffusionnft/page.md",
        "children": []
      },
      {
        "slug": "claim-3-three-new-policy-gradient-objectives-epg-pepg-par",
        "title": "Claim 3: Three new policy-gradient objectives — EPG, PEPG, PAR",
        "file": "pages/claim-3-three-new-policy-gradient-objectives-epg-pepg-par/page.md",
        "children": []
      },
      {
        "slug": "claim-4-forward-elbo-estimators-outperform-backward-trajectory-based-estimators",
        "title": "Claim 4: Forward ELBO estimators outperform backward trajectory-based estimators",
        "file": "pages/claim-4-forward-elbo-estimators-outperform-backward-trajectory-based-estimators/page.md",
        "children": []
      },
      {
        "slug": "claim-5-epg-vs-pepg-vs-par-performance-differences-are-minor-under-elbo",
        "title": "Claim 5: EPG vs PEPG vs PAR performance differences are minor under ELBO",
        "file": "pages/claim-5-epg-vs-pepg-vs-par-performance-differences-are-minor-under-elbo/page.md",
        "children": []
      },
      {
        "slug": "claim-6-evaluation-spans-geneval-ocr-drawbench-with-pickscore-clipscore-hpsv2-1",
        "title": "Claim 6: Evaluation spans GenEval, OCR, DrawBench with PickScore, CLIPScore, HPSv2.1",
        "file": "pages/claim-6-evaluation-spans-geneval-ocr-drawbench-with-pickscore-clipscore-hpsv2-1/page.md",
        "children": []
      },
      {
        "slug": "conclusion",
        "title": "Conclusion",
        "file": "pages/conclusion/page.md",
        "children": []
      }
    ]
  },
  "traces": [],
  "workspace": {
    "file": "workspace.json",
    "file_count": 0,
    "total_size": 0,
    "bucket_id": null
  },
  "agent_view_tokens": 9770,
  "trace_view_tokens": 36994,
  "workspace_view_tokens": 8,
  "revision": "a2e6418303d31b9d2b1a",
  "traces_ref": {
    "repo_id": "Firemedic15/repro-rl-design-space-diffusion-models-likelihood-estimation-traces",
    "repo_type": "dataset",
    "repo_url": "https://huggingface.co/datasets/Firemedic15/repro-rl-design-space-diffusion-models-likelihood-estimation-traces",
    "private": true
  },
  "trace_dataset": "https://huggingface.co/datasets/Firemedic15/repro-rl-design-space-diffusion-models-likelihood-estimation-traces"
}