Search papers, labs, and topics across Lattice.
Fudan University
1
0
2
Naive RL in recommender systems suffers from biased gradients that favor longer paths, but ProRL fixes this with a novel reward centering and advantage estimation scheme.
