solver.press

← All replicated papers

Inconclusive under our protocolAuthors reviewedOct 27, 2026

MetaLearnNCA: Few-Shot Offline Meta-Learning via Interacting Neural Cellular Automata

Guichard and Nichele · arXiv 2610.08479

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

Trained on Omniglot, the model reaches 89.71% 10-shot accuracy on MNIST against 79.17% for ProtoNet, and 51.26% on KMNIST against 36.23% for first-order MAML.

What we ran and what came back

All 42 checkpoints the authors released, evaluated with their own code, against baselines retrained from scratch (5 seeds each). Out-of-distribution KMNIST: +11.3 points over first-order MAML (bound +7, met). MNIST: +6.3 points over ProtoNet, interval +4.6 to +8.1; the pre-set bar of +5 was not confirmed, so the verdict under our protocol is Inconclusive, though the interval supports an advantage. The paper's single checkpoint reproduces (90.07% against 89.71% on MNIST) and sits near the top of the 42. The model itself was not retrained, so the training recipe is untested.

What this result supports

On the released checkpoints the NCA is clearly better than first-order MAML on OOD KMNIST and moderately better than ProtoNet on MNIST on average.

What it does not support

That the Table 1 gaps (+10.5 / +15.0) are typical of the method (seed 0 is near the top of the 42); that the training recipe reproduces the checkpoints.

Record

Protocol, deposited before the run: 10.5281/zenodo.23610619

Report and data: 10.5281/zenodo.23689583

The authors reviewed this result before it 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.