Gemma 4 E2B 在单块 Google TPU 上完成多轮智能体 RL 训练
单块 TPU 就能把一个 9.4% 胜率的小模型练到 95%,用的是 GRPO 多轮智能体 RL。配方全开源,想自己跑 agent 强化学习又没集群的开发者可以直接照着复现,成本门槛比想象中低很多。
Shubham Saboo 演示了在单块 Google TPU 上对 Gemma 4 E2B 运行多轮智能体强化学习。
任务是一个需要工具调用的棋盘游戏,模型初始胜率仅 9.4%。
经过 100 步 GRPO 训练后,胜率提升到 95%。完整训练配方 100% 开源,包含环境和奖励设置,开发者可以直接复现这套单卡多轮 RL 流程。
Crazyy...multi-turn agentic RL on a single Google TPU.
Gemma 4 E2B started at a 9.4% win rate on a tool-calling board game. After 100 steps of GRPO, it hit 95%.
Full recipe is 100% Opensource. https://t.co/vVTfeKnVzc
This report is categorized under AITOP's "How-to" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Gemma, TPU, GRPO, agents.
Source attribution: curated from Shubham Saboo (@Saboo_Shubham_).