CoolFace
Apppublic

SabaPivot/repro-belief-propagation-converges-to-gaussian-distributions-in-sparsely-connected-factor-graphs

sourceHugging Faceupdated 2mo agoView on Hugging Face
0likes
App README

FaGn04tvzq — Full six-claim reproduction

Direct clean-room evidence for all six claim groups in arXiv:2601.21935. The final Hugging Face cpu-upgrade run reports 6/6 aligned: 1,024-bin, 101-seed chain/tree/loopy experiments; an analytic R*=6.016848 boundary; 31/31 positive paired degree slopes; and an official 30,000-variable Middlebury Cones BP/GBP comparison.

Cones BP MAP MSE is 21.781727±0.003305 versus GBP 20.096043±0 (7.739% gap). Low-contrast mean KL is 0.012508, versus 0.215593 at edges.

The logbook explicitly reports the divergent finite-grid Figure-4c sweep and all clean-room substitutions. See artifacts/report.md for the illustrated technical article and artifacts/bp_gaussian_tutorial.py for the validated marimo tutorial.

Published with Trackio.