Weak-to-Strong Generalization via Direct On-Policy Distillation

  • 类型:arxiv
  • 标识:2607.05394
  • 链接:https://arxiv.org/abs/2607.05394
  • 主分类:engineering
  • 形态:method
  • 被引:5
  • 被引来源:Semantic Scholar
  • S2被引:5
  • OpenAlex被引:0
  • 影响力被引:0
  • TLDR:Direct On-Policy Distillation (Direct-OPD) is proposed, which transfers the teacher's RL-induced policy shift instead of running sparse-reward RL on the target model and consistently leverages weaker teachers to improve stronger target models.
  • OpenAlex ID:W7167585196
  • OpenAlex DOI:10.48550/arxiv.2607.05394
  • DOI:10.48550/arxiv.2607.05394
  • DOI来源:OpenAlex
  • 开放获取:green
  • 开放获取链接:https://doi.org/10.48550/arxiv.2607.05394
  • OpenAlex更新:2026-07-19
  • 待LLM分类:否
  • 标题中文:通过直接在线策略蒸馏实现弱到强泛化
  • TLDR中文:提出 Direct On-Policy Distillation(Direct-OPD),该方法迁移教师模型由 RL 引起的策略偏移,而非在目标模型上运行稀疏奖励 RL,并一致地利用更弱的教师模型来提升更强的目标模型
  • 来源文件
  • /inbox/tom/_candidates/2026-07-14-agent-rag-longcontext-candidates.json
  • [S2 enrich]
  • [OpenAlex backfill]