Error Diffusion Replaces Latent Weights: Training Binary-Weight Language Models on a Few Bits of Dithered State
Bonin · arXiv 2610.34566
The outcome is the verdict on the one claim below, under the protocol we fixed in advance. It is not a verdict on the paper as a whole.
Claim tested
On the width-192 binary MLP benchmark a 4-bit error-diffusion buffer with a grid level at the flip threshold matches an 8-bit buffer and stays near the latent-weight reference with 5 vs 97 bits of state; the naive clamp is worse.
What we ran and what came back
The paper's width-192 MLP benchmark (binary weights and activations), run with the authors' released code unchanged on 30 seeds the authors did not use, on CPU; their own seeds reproduced their logged means. A 4-bit error-diffusion buffer with a grid level at the flip threshold matched the 8-bit buffer (difference +0.30 points, interval -0.19 to +0.80, bound plus or minus 1.5) and was 0.4 points below the latent-weight reference (interval -0.84 to +0.05, bound -2.5) at 5 bits of state per weight against 97. The naive clamp was 3.5 points worse at 4 bits (interval -4.09 to -2.89) and the aligned 2-bit grid 19.3 points better than the naive one. Only this benchmark was tested: not the transformer results, scaling ladders, deploy-then-adapt experiments or CPU kernel, and the task is one frozen teacher-student instance. The author was sent the result at the same time as this entry and has not yet replied.
What this result supports
The width-192 MLP-benchmark relations hold in new seeds with the released code.
What it does not support
Transformer, scaling-ladder, deploy-then-adapt or CPU-memory claims; other tasks.
Record
Protocol, deposited before the run: 10.5281/zenodo.23799146
The authors were sent this result. They have not reviewed this listing.
How the protocol works is on the method page. If you are an author of this paper and something here is wrong, write to customers@aegismind.app and we will correct it on this page.