AI Papers Daily

#Window-Level Optimization

1 篇

Performance-Driven Policy Optimization for Speculative Decoding with Adaptive Windowing

基于性能驱动的自适应窗口策略优化用于投机解码

🧭 Training ★ 深度解读

提出PPOW框架，将草稿模型训练转为窗口级强化学习，结合性能奖励与自适应窗口采样提升投机解码效率

Training | 2026-05-14 | cs.CL | score 9/10

Speculative DecodingReinforcement LearningLLM InferenceWindow-Level OptimizationDrafter Training