https://sunyuhan19981208.github.io/Yuhan 的研究笔记：LLM、强化学习与Agent。zh-CNhttps://sunyuhan19981208.github.io/posts/multiagent-opd/https://sunyuhan19981208.github.io/posts/multiagent-opd/Fri, 16 Oct 2026 00:00:00 +0800Train coordination and answering from outcomes, while a task-trained teacher guides the search tokens the student actually generates.https://sunyuhan19981208.github.io/posts/harness-aware-training/https://sunyuhan19981208.github.io/posts/harness-aware-training/Sun, 16 Aug 2026 00:00:00 +0800Tools, rules, and prompts change. Training should not assume one fixed runtime.https://sunyuhan19981208.github.io/posts/agenticrl/https://sunyuhan19981208.github.io/posts/agenticrl/Mon, 13 Jul 2026 00:00:00 +0800从 Workflow 升级到 Agentic 架构，介绍 AgentTuning、回复模型 RLVR、端到端 Agentic RL 与多智能体协同强化学习的实践。https://sunyuhan19981208.github.io/posts/livethinking/https://sunyuhan19981208.github.io/posts/livethinking/Thu, 09 Oct 2025 00:00:00 +0800Distill the capability, then use reinforcement learning to remove unnecessary thinking.https://sunyuhan19981208.github.io/posts/llm-barrage-interaction/https://sunyuhan19981208.github.io/posts/llm-barrage-interaction/Wed, 25 Jun 2025 00:00:00 +0800从业务背景和互动架构出发，介绍多轮意图识别、实体解析、商品检索与互动回复生成的实践。https://sunyuhan19981208.github.io/posts/llm-distillation-and-reasoning-compression/https://sunyuhan19981208.github.io/posts/llm-distillation-and-reasoning-compression/Fri, 20 Jun 2025 00:00:00 +0800从 DeepSeek-R1 蒸馏到推理路径压缩，介绍 long2short DPO、质量与长度双偏好 DP-DPO 及泛化性优化实验。https://sunyuhan19981208.github.io/posts/controlpe/https://sunyuhan19981208.github.io/posts/controlpe/Thu, 16 Nov 2023 00:00:00 +0800Distill a prompt’s effect into LoRA, then explore the space between applying it fully and leaving it out.