Smaller Abstract State Spaces Enable Cross-Scale Generalization in Reinforcement Learning

In three linesTheoretical paper on out-of-distribution generalization in RL. Authors extend state abstraction framework to POMDPs and introduce successor-weighted model reduction. They prove smaller abstract state spaces improve generalization to more complex tasks.

Summary generated by Claude — human-verified