title: Sinkhorn Diffusion — Six-Claim CPU Reproduction
emoji: 📐
colorFrom: blue
colorTo: green
sdk: static
app_file: index.html
tags:
- icml2026-repro
- paper-LJdacnMXkr
Six Sinkhorn diffusion certificates
This frozen, unpublished package audits all six operative anchored claims for Sinkhorn Normalization of Diffusion Kernels (LJdacnMXkr, arXiv 2507.06161v2). The conservative prerelease forecast is 12/12; that is not an official score.
The evidence combines finite matrix identities for symmetric diagonal scaling, Gaussian/exponential refinement grids, an explicit 0.1% iteration audit, joint diffusion-axiom checks, three irregular data modalities, and a source-locked Stanford Armadillo spectral comparison. Eight destructive controls are executed, not merely described.
Run locally with CPU-only settings:
CUDA_VISIBLE_DEVICES='' PYTORCH_ENABLE_MPS_FALLBACK=0 \
OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 \
python reproduction/run_tests.py
Future artifact Bucket: https://huggingface.co/buckets/neonforestmist/sinkhorn-normalization-diffusion-kernels-repro-artifacts
The USE_IT_LOCK forbids publication or consumption until the user explicitly says USE IT.
Xet Storage Details
- Size:
- 1.24 kB
- Xet hash:
- 14701b8d0e094a9e57cd3c13e864527837b6b68a6f8da1a262d9a70acc590af7
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.