Reproduce “Continual-ARC: Measuring Skill Acquisition over a Working Lifetime” 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
Daniel Lenton
Continual-ARC tests whether systems learn across recurring ARC puzzles with fresh inputs and generalize to unseen tasks. It compares memory-based harnesses with DAgger, which trains a low-rank adapter to imitate a guided teacher. On Qwen3.5-9B’s ARC-AGI-1 tests, DAgger saves 0.57 of reset feedback-budget spend on seen tasks and 0.29 on held-out tasks; this setup does not isolate weight updates.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.