ThinkingBox-Training turns database checks into RL rewards
Microsoft and Hugging Face report that ThinkingBox-Training is now available on GitHub, using executable checks of final database states as reinforcement-learning rewards for agents. In their reported experiment, Qwen3.8-27B improved from 51.70% to 60.78% pass@1 after 50 GRPO steps, while tasks completed correctly in all 20 attempts increased from 38 to 107. However, pass@20 fell from 89.35% to 87.97%, and the authors note that shared benchmark elements limit the transfer claim.
The release provides an open implementation for training stateful agents with verifiable database-state rewards, while the mixed benchmark results show why repeatable success and broad task coverage should be evaluated separately.
Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox
Published October 22, 2026
#Microsoft#Hugging Face#ThinkingBox-Training#ThinkingBox-Bench#Qwen3.8-27B#Microsoft Copilot Studio#Toloka