List your company

Agent Tool-calling on Pokémon Red (20 held-out transitions)

50Action Applicability RateGemma-4 31B
-211.52538.5May 11, 2026

Evaluation results

MethodLinks
Gemma-4 31BStage=SFT2026.055050756544
Gemma-4 26BStage=Offline GRPO (heuristic)2026.055050655044
Gemma-4 26BStage=Offline GRPO (Gemini oracle)2026.055050553040
Gemma-4 26BStage=Base2026.05255332523
Qwen3.5 35BStage=Base2026.050————