solver.press

← All replicated papers

Did not replicate under our protocolOct 30, 2026

CNet: A Complex-Valued Deep Learning Framework with Wirtinger Autodifferentiation and FFT--Hadamard Convolution

Crasmaru · arXiv 2610.08592

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

A fully complex FNet-style causal LM matches a param-matched real-valued FNet in under half the steps on tiny-shakespeare.

What we ran and what came back

The authors' released code, run on two T4 GPUs: a complex-valued FNet-style causal language model on tiny-shakespeare, 15,000 steps, the example's default settings, one seed per run. The paper's curve ends near 1.75 and reaches about 2.03 by step 5,000. Our smoothed held-out loss ended at 1.926 (the commit we pinned, original start), 1.924 (the authors' current head, old start) and 1.903 (current head, new default start), and first reached 2.03 at step 9,000, 8,500 and 8,500. Our pre-registered gate (held-out loss at most 1.80 by step 5,500) failed in all three runs. The logged training loss ended at 1.76 to 1.78, close to the paper's 1.75, and the paper does not say whether its curve is training or held-out loss. All controls passed. The paper's claim about matching a real-valued model in under half the steps, the Born-attention model and the real-valued baseline were not tested. One seed per run, untuned learning rate, two ranks of 16; the result applies to these commits.

What this result supports

At the paper's model size and the example's settings the released code's held-out loss ends near 1.91-1.94.

What it does not support

Anything about the Born-attention model, the real-valued baseline claim, or whether the paper's curve is training loss.

Record

Protocol, deposited before the run: 10.5281/zenodo.23332645, 10.5281/zenodo.23764212

The authors were sent this result. They have not reviewed this listing. The authors were emailed on 12 and 26 October, and the result and a two-page report were sent on 30 October 2026. No reply had been received when this was listed.

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.