SEED: Self-Evolving RL for Agents
heghbalz · x · 2026-07-19
这条转发介绍了一篇名为 SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning 的论文。
核心思路是：把每次完成的 agent trajectory 变成一个“hindsight skill”，再用同一套持续演化的 policy 重新给采样动作打分，于是同一模型同时承担两件事：
- 用稀疏的最终结果做 RL 训练
- 用密集的 token 级信号做蒸馏
配图展示了它在 ALFWorld、Search-based QA、WebShop 三个基准上的结果，相比 vanilla、GRPO、skill-distillation 等方法，SEED 在多个指标上更高，尤其在 ALFWorld 平均分达到 91.8，Search-based QA 平均 45.7，WebShop 成功率 78.9。
Related event: SEED: Self-Evolving Distillation for Agentic RL(2 posts)→
More from coding & agent
- Soft Clamp cuts tool-call overuse in multi-teacher distillation, from 13.7% to 9.0% — antgroup · 2026-07-21
- SpecJudge runs locally on Ollama to pick the right-sized AI model for your project — jokiruiz · 2026-07-21
- A developer maps out six design rules for CLIs that humans and AI agents can both use — yujiezha · 2026-07-21
- A coding-agent skill that forces ADHD-friendly, answer-first output — ayghri · 2026-07-21
- A set of agent skills for CAD, robotics, and hardware design — earthtojake · 2026-07-21
- Outlines keeps LLMs on-rails with structured outputs — dottxt-ai · 2026-07-21