Search papers, labs, and topics across Lattice.
Fudan University
2
0
4
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.
Stop training LLMs on lucky guesses: this new RL method uses the model's own in-context learning ability to identify and upweight high-quality reasoning traces, leading to better performance.
