Search papers, labs, and topics across Lattice.
University of Colorado Boulder
3
0
8
Seemingly minor restrictions on generator access during post-training can create exponential gaps in performance, suggesting that the interface between learner and generator is a critical, often overlooked, factor.
Standard data attribution methods break down in adaptive learning scenarios, but this work identifies conditions under which you *can* still recover meaningful attributions from logged data.
Forget static arms: this new bandit framework tackles dynamically changing action sets, revealing the fundamental cost of exploration under local-move constraints.
