← News

ResearchSources: arXiv · 19 Nov 2026

Lamarck strengthens expert adapters for replay-free domain-incremental learning

A training-free router plus inherited, token-correcting adapters lift a frozen ViT approach to the top of a deepfake benchmark, its authors report.

In plain words

A model must keep learning new visual domains without storing old images. Lamarck trains a small expert for each domain and copies each expert from the previous one, so mistakes hurt less. The authors say this beats twelve baselines on a deepfake benchmark.

What was announced

The arXiv abstract describes replay-free domain-incremental learning, where a model trains on the same classes across successive domains without keeping earlier images. Each domain gets its own expert: a small set of parameters plus a classification head, added to a frozen pre-trained ViT. Because test images carry no domain label, a router picks one expert per image. The authors find a training-free Mahalanobis router costs under 2.7 points of final accuracy versus always picking the own-domain expert, so they conclude the experts, not the router, limit accuracy.

Their system, Lamarck, therefore keeps the training-free router and strengthens the experts. Its adapters are small trainable layers that correct every image token after each deep ViT block, adding 4.4 points of final accuracy on the deepfake benchmark CDDB-Hard. Each expert is also initialized from its predecessor at no extra cost, which the authors say limits the damage when an image is misrouted to a later expert, since domains share classes. Under one protocol against twelve frozen-backbone baselines, Lamarck ranks first on CDDB-Hard by 5.2 points, second on DomainNet and OfficeHome, and highest in mean accuracy. These figures come from the paper's own abstract and have not been independently confirmed.

What it does not show

  • Evidence is limited to the paper's abstract; the full method, experiments and protocol details are not covered, and the reported rankings are the authors' own claims without independent confirmation.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.