用信任区域融合教师行为，提升学生初始推理质量。
Trust-Region Behavior Blending for On-Policy Distillation
- 在信任区域中用教师相近行为替代学生早期差策略
- 两个数学推理任务上表现优于现有方法
- 适合需要稳定训练初期的强化学习迁移场景
在线策略蒸馏（OPD）让学生基于自身策略采样前缀进行训练，以匹配更强的教师模型，解决了离线蒸馏中的前缀不匹配问题。但学生早期采样可能质量差，导致教师监督作用于低质量前缀。本文提出信任区域行为融合（TRB），一种预热方法：在学生中心的KL信任区域内，用最接近教师行为的策略替换早期采样策略，同时保持每前缀反向KL损失不变。KL预算随训练逐步衰减至零，训练后期回归纯学生采样。在两个数学推理蒸馏设置中，TRB在平均性能上超越所有对比方法。
原文摘要 · Abstract (English)
On-policy distillation (OPD) trains a student on prefixes sampled from its own policy while matching a stronger teacher. This addresses the prefix mismatch of offline distillation, but early student rollouts can still be poor, placing teacher supervision on weak or low-quality prefixes. We propose Trust-Region behavior Blending (TRB), a warmup method that replaces the early rollout policy with the closest-to-teacher behavior policy inside a student-centered KL trust region, while keeping the per-prefix reverse-KL OPD loss unchanged. The KL budget is annealed to zero, so training returns to pure student rollouts after warmup. Across two math-reasoning distillation settings, TRB attains the strongest average among the compared methods.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。