Reproduce “BISECT: Bilevel Spectral Decomposition for Low-Rank Continual Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Tom Saliencro, Wen-Hsuan Liu, Rohan Desai, Amara Okonkwo, Maya Lindqvist

BISECT addresses exemplar-free class-incremental learning by adapting a vision model to new classes without replaying old examples or using task IDs at test time. It splits a fixed LoRA rank budget between shared and relatively novel feature directions, then gates write-back to shared directions. On 20-session ImageNet-A, it reaches 56.12% final accuracy versus InfLoRA’s 46.95% under the reported protocol.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.