| { | |
| "claim_title_fix": "2026-07-22T17:10:00Z", | |
| "schema_version": 1, | |
| "title": "Repro: High-accuracy sampling for diffusion models and log-concave distributions", | |
| "emoji": "🎯", | |
| "space_id": "snaykey/repro-high-accuracy-sampling", | |
| "paper": { | |
| "arxiv_id": "2602.01338", | |
| "openreview_id": "GW3umRqsZZ" | |
| }, | |
| "tags": [ | |
| "icml2026-repro", | |
| "paper-GW3umRqsZZ" | |
| ], | |
| "updated_at": "2026-07-19T11:37:20+00:00", | |
| "root": { | |
| "slug": "index", | |
| "title": "Repro: High-accuracy sampling for diffusion models and log-concave distributions", | |
| "file": "pages/index.md", | |
| "children": [ | |
| { | |
| "slug": "claim-1-diffusion-sampler-polylog-1-delta-steps-theorem-4-3", | |
| "title": "The paper's First-Order Rejection Sampling (FORS) meta-algorithm (Theorem 3.1) produces samples with error δ using sample complexity bounded by 3Be^(2B)log(2/δ) with probability 1-δ (Theorem 3.1).", | |
| "file": "pages/claim-1-diffusion-sampler-polylog-1-delta-steps-theorem-4-3/page.md", | |
| "children": [] | |
| }, | |
| { | |
| "slug": "claim-2-complexity-o-tilde-d-polylog-1-delta-theorem-4-3", | |
| "title": "Under only a finite second-moment assumption (minimal assumptions), the diffusion sampler achieves query complexity O(d·log²(1/δ) + log³(1/δ)), giving polylog(1/δ) dependence rather than the poly(1/δ) of prior work (Theorem 4.1, Section 4).", | |
| "file": "pages/claim-2-complexity-o-tilde-d-polylog-1-delta-theorem-4-3/page.md", | |
| "children": [] | |
| }, | |
| { | |
| "slug": "claim-3-intrinsic-dimension-d-star-complexity-corollary-4-4", | |
| "title": "Under a non-uniform Lipschitz condition on the score (Assumption 4.3), a DDPM-like sampler achieves total-variation error controlled via chi-squared divergence with complexity O(√(dL_δ log(d/δ))·log(d/δ) + L_δ log²(d/δ)) (Theorem 4.4).", | |
| "file": "pages/claim-3-intrinsic-dimension-d-star-complexity-corollary-4-4/page.md", | |
| "children": [] | |
| }, | |
| { | |
| "slug": "claim-4-non-uniform-lipschitz-refinement-theorem-4-9", | |
| "title": "For distributions with low intrinsic dimension d★, an adaptive-step-size method attains complexity O(d★·log²((d+M₂²)/δ²)), replacing the ambient dimension d with d★ (Theorem 4.6).", | |
| "file": "pages/claim-4-non-uniform-lipschitz-refinement-theorem-4-9/page.md", | |
| "children": [] | |
| }, | |
| { | |
| "slug": "claim-5-log-concave-sampling-via-gradient-queries-section-5", | |
| "title": "Section 5 extends the FORS framework to sample from general log-concave distributions using only gradient evaluations (no density evaluations), giving the first polylog(1/δ) sampler in this setting (Section 5).", | |
| "file": "pages/claim-5-log-concave-sampling-via-gradient-queries-section-5/page.md", | |
| "children": [] | |
| }, | |
| { | |
| "slug": "conclusion", | |
| "title": "Conclusion", | |
| "file": "pages/conclusion/page.md", | |
| "children": [] | |
| } | |
| ] | |
| }, | |
| "agent_view_tokens": 8606, | |
| "revision": "1784461040769985400" | |
| } |