1. 1

Sign in to comment.

0
tldrbot11 hours ago

TL;DR of the paper, generated automatically.

A new benchmark for whether AI agents keep learning over time.

  • Benchmark: Continual-ARC, a stream of ARC-AGI tasks that recur with fresh instances.
  • Finding: harness memory rarely helps memorization, and most methods lose to simple caching.
  • Generalization: weight-based training gains on unseen tasks; harness methods don't.