Qwen
Author of "Qwen Explains Why Minimizing Training-Inference Gaps and Policy Staleness Improv" on arXiv
How media typically covers Kai Dang
Kai Dang as author
Minimizing training-inference discrepancy and policy staleness enables reliable token-level surrogate optimization in RL with LLMs, with importance sampling correction and routing replay being essential for stable training at scale.
“Author of "Qwen Explains Why Minimizing Training-Inference Gaps and Policy Staleness Improv" on arXiv”