Huazheng Wang

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

SPD achieves unprecedented decoding speed by processing multiple tokens in parallel while eliminating latency bubbles, setting a new standard for LLM inference efficiency.

Multi-agent RL can boost LLM accuracy, but the benefits hinge on the intricate balance of workflow, task, and model scale rather than just policy-sharing strategies.