tracked author

Idan Shenfeld

SDFT first and corresponding author, SDPO coauthor, and MIT PhD student advised by Pulkit Agrawal. His public work connects on-policy learning, continual learning, language-model post-training, and robotics.

2 archived notes X: high Homepage GitHub X

Related Notes

按论文归档时间排序，展示该作者在本站已经出现的材料。

归档 Jul 28, 2026

Reinforcement Learning via Self Distillation

SDPO 在学生在线轨迹上用环境反馈条件化的 EMA 同模型重算下一词分布，将 Top K 分布差转成 logit 级信用；Qwen3 8B 在 LiveCodeBench v6 上达到 48.8%，高于强 GRPO 的 41.2%，但依赖模型回溯能力与反馈质量。

待审阅 2601.20802-reinforcement-learning-via-self-distillation On-Policy Distillation Credit Assignment RL Algorithm

归档 Jul 17, 2026

Self Distillation Enables Continual Learning

让 student 在自身轨迹上接受 query specific demonstration conditioned EMA 同模型的 token distribution，在提升新任务准确率时显著减少旧能力遗忘。

待审阅 2601.19897-self-distillation-continual-learning On-Policy Distillation Training Stability