Understand this story
In a tutorial, Google demonstrated a multi-turn tool-calling agent using Gemma 4 E2B trained with Tunix's agentic reinforcement learning workflow. After 100 steps of agentic GRPO training on a single Cloud TPU v6e-8, the agent's win rate on held-out Plane Strike boards rose from 9.4% to 95%, while its average number of strikes per game fell from 22.44 to about 16. The tutorial also covers the task environment, Gemma 4's native tool calling, conversational agents, and the asynchronous vllm-tpu training workflow, and provides the complete code on GitHub. There are currently no other evidence-backed developments.
AI-generated from public reports
Latest development · Report
Google Tutorial: Training a Multi-Turn Tool-Calling Agent with Tunix, Gemma 4, and TPUsDevelopments
1 confirmed developments
Gemma 4 E2B achieves a 95% win rate after agentic GRPO training
Source:Google Developers Blog
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 19
Google Developers BlogFirst-party
Google Tutorial: Training a Multi-Turn Tool-Calling Agent with Tunix, Gemma 4, and TPUs
Google demonstrates how to use Tunix's agentic reinforcement learning workflow on a single Cloud TPU v6e-8 to train a multi-turn tool-calling agent based on Gemma 4 E2B. After 100 steps of agentic GRPO training, the win rate on held-out Plane Strike boards rose from 9.4% to 95%, while the average number of strikes per game fell from 22.44 to about 16. The article breaks down the task environment, Gemma 4's native tool calling, the conversational agent, and the asynchronous vllm-tpu training workflow, and provides the complete code on GitHub.
AI Trends