Multi-turn agentic RL on one TPU lifts Gemma 4 E2B from 9.4% to 95% win rate
Saboo_Shubham_ · x · 2026-11-21
A fully open-source recipe trains Gemma 4 E2B (thinking mode on) with GRPO on a single Google TPU, playing a tool-calling board game in multi-turn conversation.
After 100 training steps, win rate climbed from 9.4% to 95%, with shorter dialogues, fewer repeated/invalid moves, and near-perfect tool-call rates. The author notes turn-level credit assignment could push results further. Built with Tunix.
More from coding & agent
- KV Cache Rules Everything: agent context re-reads cost you a lot, providers almost nothing — AccBalanced · 2026-11-21
- StepFun unveils Step 5 Preview at SF event focused on agent reliability and cost — StepFun_ai · 2026-11-21
- AI-made "game in a week" hype draws fire: it's a rough demo, not a game — AIandDesign · 2026-11-21
- Swapping RAG ranking for Jev lifted recruiter match precision from 44% to 92% — Select_Condition2805 · 2026-11-21
- What a Per-Agent Audit Trail Needs to Capture—and Where It Goes Incomplete — BlitzBallBruiser · 2026-11-21
- uivoid turns any API into a hosted, authenticated MCP server with zero code — Necessary_Client_114 · 2026-11-21