Quanquan Gu

KL-regularization can still yield high-probability guarantees even when models are misspecified, challenging the assumption that realizability is necessary for effective learning.

Self-distillation can significantly enhance reinforcement learning stability and performance, outperforming traditional methods in sparse-reward environments.

Gated Delta Networks can achieve stable learning-rate transfer across model widths, a breakthrough that significantly enhances training efficiency for large language models.

KL-regularization in multi-armed bandits provably achieves near-optimal regret, scaling linearly with the number of arms, a significant improvement over classical results.