ICML 2026 evidence package

DiffRes: traceable 100-ID Gaussian-mixture study

Historic stored output · static package · no numerical execution

A local study wrapper using released implementation, stored numerical random keys and pinned source config recorded a traceable 100-ID Gaussian-mixture result; a separate validator recomputed raw-record aggregates, paired intervals, and source hashes.

Interpretation boundary

This is not an exact reproduction, independent replication, complete Table 1 or Table 2 reproduction, publication score, or broad outperformance claim. It is a stored, traceable comparison against the multinomial baseline only. This static package does not verify a source checkout or numerical execution, and it does not establish source-hash identity, execution provenance, paper-level results, or significance.

Recorded configuration

  • Released zgbkdlm/diffres source at 767effe3e755067eb8a04422597fbf37eb8ab754.
  • Stored Monte Carlo IDs 0–99; 10,000 particles; 1,000 sliced-Wasserstein projections.
  • Probability-flow ODE, T=3, 128 steps, Jentzen--Kloeden integrator.
  • Diffusion resampling and multinomial baseline; no other Table 1 baselines.

Historical stored aggregates

Metric (lower is better) Diffusion resampling Multinomial baseline
Sliced Wasserstein L1, mean ± population SD 0.0807686 ± 0.0212899 0.0824334 ± 0.0250323
Squared posterior-mean residual L2, mean ± population SD 0.0373908 ± 0.0299358 0.0378330 ± 0.0442877

The recorded paired 95% t intervals cross zero for both metrics. They are not evidence of a statistically significant improvement.

Inspect the evidence

Selection and safety

The package contains only static presentation files, two historical JSON records, a non-networked audit script, and licensing/provenance notes. It deliberately excludes the upstream source tree, virtual environments, caches, runners, model artefacts, partial studies, and LGSSM material.