EVENT DOSSIER
On September 7, 2026, the Salesforce AI Research team proposed a new large-scale model training method called RISE (Recursive Improvement via Self-Extrapolating Policy Distillation). This method aims to address the issue of limited quality of teacher models in the post-training phase of language models, without relying on external teachers or privileged data. RISE utilizes the model’s own reinforcement learning validation trajectory, converting sparse parameter updates into dense token-level supervision signals through self-extrapolation techniques. Specifically, the algorithm extrapolates based on the displacement of the current checkpoint and the tail anchor, guides the direction of results with combined rewards, and refines token decisions to form a recursive improvement loop. Experiments were conducted in areas such as mathematical reasoning, multi-field STEM, code generation, and multi-round agent tasks, and results showed that RISE outperformed both training using only RLVR and policy-based self-distillation methods across all test settings.
Salesforce AI Research RISE 提出基于自外推策略蒸馏的递归改进大模型方法
The Salesforce AI Research team proposed RISE (Recursive Improvement via Self-Extrapolating Policy Distillation). This …
2 reportsarXiv:2609.05295v1 提出名为 RISE（Recursive Improvement via Self-Extrapolating Policy Distillation）的新方法，旨在通过自外推策略蒸馏解决语言模型后训练中的教师质量瓶颈。该方法利用模型自身的 RLVR 训练轨迹构建合成教师，通过外推当前检查点与尾部锚点之间的位移，将稀疏的参数更新转化为稠密的 token 级目标，无需外部模型或特权条件。RISE 将 RLVR 与 OPD 结合为互补循环：结果奖励引导外推方向，外推教师细化 token 决策，且教师随学生迭代刷新形成递归改进机制。实验涵盖数学推理、多领域 STEM、代码生成及多轮智能体任务，结果显示 RISE 在所有设置下均优于仅 RLVR 训练和 OPD 自蒸馏。
The Salesforce AI Research team proposed RISE (Recursive Improvement via Self-Extrapolating Policy Distillation). This method recursively generates dense token-level supervision from the model’s own reinforcement learning trajectory using self-extrapolation techniques, without the need for external teachers. It utilizes the displacement between the current checkpoint and the tail anchor for extrapolation, transforming sparse result-induced parameter updates into dense targets. Experiments covered mathematical reasoning, multi-field STEM, code generation, and multi-round agent tasks, and results showed that RISE outperformed both RLVR-only training and policy-based self-distillation methods in all settings.