Pre-registration of a replication of the width-192 MLP benchmark claims of 'Error Diffusion Replaces Latent Weights' (arXiv 2610.34566)
Year: 2026
Type: article
Abstract: Frozen pre-registration (30 Oct 2026) of a private replication of four quantitative claims of arXiv 2610.34566 on the authors' released teacher-student MLP benchmark (repo pierreolivierbonin/error-diffusion-binary-weight-lms, commit 436b6aa6): a 4-bit error-diffusion buffer with a grid level at the flip threshold matches an 8-bit buffer, stays within a stated margin of the latent-weight reference, the naive clamp is worse, and the 2-bit aligned grid recovers most of the loss. Confirmatory seeds 200-229, reproduction seeds 0-4, CPU. Contains the decision code, harness, power simulation, planted-effect table and the run kernel. Deposited before the main runs. Outcome wording: Replicated / Did not replicate / Inconclusive under protocol X.
Author: Goodman, John
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7224750285