Google uses Tunix and Gemma 4 to improve the win rate of LLM agents
The story so far
Google uses Tunix and Gemma 4 trained on Google TPUs to train multi-round tool using LLM agents, increasing the win rate in Plane Strike game from 9.4% to 95%, and shortening the average game length to about 16 steps. The agent adapts strategies through multi-round reasoning, tool calls, and environmental feedback, and the code is open source. The agent adapts strategies through multi-round reasoning, tool calls, and environmental feedback, and the code is open source.
AI-written from the reports · Updated 6 hours ago
Coverage timeline
Follow the coverage to see every side of the story.
- Google Developers BlogMulti-Turn Agentic Reinforcement Learning with Tunix and Gemma 4 on Google TPUs
The multi-round tool-using LLM agent trained on Google TPUs with Tunix and Gemma 4 increased its win rate in Plane Strike from 9.4% to 95%, with the average game length shortened to about 16 steps. The agent adapts strategies through multi-round reasoning, tool calls, and environmental feedback, and the code is open source.
Heat over time
- Now
- 5
- Peak
- 620 Nov, 02:00
- 24h change
- –
Hover or tap the chart to see each hour; use the arrow keys on a keyboard.