| title: High Accuracy Sampling Reproduction | |
| emoji: 📈 | |
| colorFrom: blue | |
| colorTo: green | |
| sdk: gradio | |
| sdk_version: 6.22.0 | |
| app_file: app.py | |
| tags: | |
| - icml2026-repro | |
| - paper-71132 | |
| # Reproduction: High-accuracy sampling for diffusion models and log-concave distributions | |
| **Paper ID:** `71132` | |
| **Upstream Revision:** `arxiv:2602.01338v2+arxiv-source:2602.01338v2` | |
| **Challenge:** ICML 2026 Agent Repro Challenge | |
| --- | |
| ## Overview | |
| This repository provides an independently executable reproduction bundle verifying key theoretical and algorithmic bounds from *High-accuracy sampling for diffusion models and log-concave distributions* (Chen et al., 2026). | |
| --- | |
| ## Claims Verified | |
| 1. **Theorem 4.3 (Polylog Step Scaling):** | |
| The diffusion sampler achieves $\delta$-error in $\tilde{O}(\text{polylog}(1/\delta))$ steps given sufficiently accurate score estimates. | |
| 2. **Corollary 4.4 (Intrinsic Dimension Reduction):** | |
| When data has intrinsic dimension $d^*$, step complexity scales as $\tilde{O}(d^* \text{polylog}(1/\delta))$. | |
| 3. **Section 5 (Log-Concave First-Order Sampler):** | |
| A first-order gradient query sampler for log-concave distributions achieves $\text{polylog}(1/\delta)$ accuracy. | |
| --- | |
| ## Evidence Generation | |
| To run the evidence generation suite and update `evidence/bundle.json`: | |
| ```bash | |
| python generate_evidence.py | |
| ``` | |