MultiAgent-OPD: hierarchical multi-agent training
Train coordination and answering from outcomes, while a task-trained teacher guides the search tokens the student actually generates.
Y'LOG / COLLECTION
Posts
7 篇文章 · 均含中英文版本
Harness-Aware Training: adapting to runtime changes
Tools, rules, and prompts change. Training should not assume one fixed runtime.
淘宝直播数字人 AgenticRL 实践：从 RLVR 到 MultiAgent RL
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。
LiveThinking: efficient reasoning for real-time QA
Distill the capability, then use reinforcement learning to remove unnecessary thinking.
淘宝直播数字人：LLM弹幕互动技术
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。
淘宝直播数字人LLM推理优化：模型蒸馏与路径压缩实践
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。
ControlPE: controlling prompt strength
Distill a prompt’s effect into LoRA, then explore the space between applying it fully and leaving it out.