Search papers, labs, and topics across Lattice.
4
0
5
Discounted occupancy-ratio realizability alone can enable robust offline policy evaluation, eliminating the need for stringent completeness assumptions.
Mult-DPO reveals that a multinomial approach can effectively align LLMs with complex user preferences, outperforming traditional pairwise methods.
Unlock $\sqrt{N}$ regret in offline policy learning, even with complex policy classes, by trading off policy and environment complexity.
Unlock asymptotically normal and semiparametrically efficient estimators in adaptive data collection by using a novel target-specific condition called "directional stability," which is weaker than previous target-agnostic conditions.
