Continual-ARC: lifelong ARC-AGI stream finds agent harness persistence rarely beats simple caching

unifyai · hf · 2026-11-16

Unify AI proposes Continual-ARC, a lifelong stream built from the three ARC-AGI editions where tasks return after dormant gaps (testing memorization) and new tasks keep arriving (testing generalization). Learners get a per-instance feedback budget to buy demonstrations, submit answers, or act inside the harness, scoring memory- and weight-based systems on one scale. Frozen-weight experiments on Qwen3.5-9B and GLM-5.3 Flash show harness persistence rarely improves memorization, most harness methods underperform simple caching, and weight-based learning generalizes better.

Original post →

More from coding & agent

coding & agent channel →