Long-form notes on machine learning, research, and building systems.
- Writing on agentic RL, post-training, tooling, and research workflows.
- Built with Hugo PaperMod and integrated into the main portfolio as a subsite.
Long-form notes on machine learning, research, and building systems.
从信息闭环、状态性、不可逆动作和决策质量度量重新定义 agentic RL,区分它与 RLM、single-turn tool use 和 verifiable multi-turn tool use 的本质边界。
围绕 Agent-World 论文,梳理开源小模型与闭源通用模型在复杂 agentic 任务上的差距、环境 scaling law 的形态以及 self-evolution 的收益结构。
按 token、rollout、group、batch 和训练过程等尺度梳理 RL post-training 中应观察的诊断指标,以及每个指标能说明和不能说明什么。