This purely theoretical paper introduces and studies new models of query learning with contrastive examples.

We use grokking to disentangle generalization from training dynamics and show that relative flatness, not neural collapse, is a necessary and more predictive indicator of generalization in deep networks.

To advance evaluation of RPOMDP policies, we (1) introduce a formalization for suitable benchmarks, (2) define a new evaluation method, and (3) lift existing POMDP value bounds to RPOMDPs.

We characterize and provide algorithms for multi-environment POMDPs.