Reproduce “Found by Search, Broken by a Shortcut: Transfer and Tampering of Searched Intrinsic Rewards” 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
Chase J. Metoyer
The authors search generic event-based intrinsic rewards for reinforcement-learning agents, using needs tied to seeing, picking up, dropping, and toggling objects. Agents train with task reward plus the searched drive. On held-out KeyCorridorS3R2, reward E exceeded random-search-selected R by 0.159 in mean training-curve area (Holm-adjusted p = 0.00015); this comparison is limited to one task, and E failed under an unseen reset shortcut.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.