wrice's picture
Publish validated 601ee343aade48dee6ca273bb589571c13ace225
b39dfd0 verified
Raw
History Blame Contribute Delete
1.3 kB
"""Gradio Space app for High-accuracy sampling reproduction."""
import json
from pathlib import Path
import gradio as gr
def load_bundle() -> dict:
bundle_path = Path(__file__).parent / "evidence" / "bundle.json"
if bundle_path.exists():
return json.loads(bundle_path.read_text())
return {"error": "Evidence bundle not found"}
bundle_data = load_bundle()
with gr.Blocks(title="High-Accuracy Diffusion Sampling Reproduction") as demo:
gr.Markdown("# Reproduction: High-accuracy sampling for diffusion models and log-concave distributions")
gr.Markdown("Paper ID: `71132` | Upstream Revision: `arxiv:2602.01338v2` | ICML 2026 Agent Repro Challenge")
with gr.Tab("Claims & Evidence"):
gr.JSON(bundle_data)
with gr.Tab("Summary"):
gr.Markdown("""
### Target Claims
1. **Theorem 4.3**: Diffusion sampler achieves $\\delta$-error in $\\tilde{O}(\\text{polylog}(1/\\delta))$ steps given accurate score estimates.
2. **Corollary 4.4**: Complexity reduces to $\\tilde{O}(d^* \\text{polylog}(1/\\delta))$ under intrinsic dimension $d^*$.
3. **Section 5**: Polylogarithmic accuracy sampler for log-concave distributions using first-order gradient queries.
""")
if __name__ == "__main__":
demo.launch()