Microsoft releases ThinkingBox code for training database agents

Microsoft has released the ThinkingBox-Training code repository on GitHub, as reported on the Hugging Face Blog. The tool evaluates the output of artificial intelligence agents by comparing the terminal state of a database against expected outcomes. It then uses those execution signals as rewards during post-training reinforcement learning on tasks outside its benchmark dataset. [1]

The Hugging Face Blog reported that applying these executable reward signals to the Qwen3.8-27B model improved its performance. Across 50 Group Relative Policy Optimization steps, the average success rate of the model rose from 51.70 percent to 60.78 percent. Tasks completed correctly on all 20 attempts increased from 38 to 107. [1]

Reinforcement learning is a core paradigm in machine learning that focuses on training agents to take actions in dynamic environments to maximize cumulative reward signals. The release follows earlier evaluation work documented by Microsoft in its previous technical report. [1][3]

The new open release allows developers to apply outcome-based reward signals to refine database manipulation agents beyond static benchmarking tasks. [1]