Microsoft and Hugging Face open-source ThinkingBox-Training for database agent RL

Microsoft and Hugging Face have released ThinkingBox-Training on GitHub, a method that converts the database grading system from their ThinkingBox benchmark into reinforcement learning rewards for post-training agents. The approach uses 'offline policy-relative difficulty' to select appropriately challenging training tasks, then applies GRPO fine-tuning with binary rewards derived from final database state checks. On Qwen3.8-27B, this raised average success from 51.70% to 60.78% and more than doubled the number of tasks completed correctly on all 20 attempts from 38 to 107.

How executable grading becomes training signal

• ThinkingBox-Bench contains 507 tasks with real backends and MCP tools; agents are graded by comparing terminal database state against expected outcomes, not by tool-call sequences
• Binary reward signal: most tasks yield a direct 1/0 reward from executable database checks, making the grader itself usable for RL
• Offline policy-relative difficulty: tasks are filtered by estimated success probability under the base policy; already-solved and hopeless tasks are excluded from training
• GRPO fine-tuning: 50 steps on Qwen3.8-27B with the selected training pool
• Tradeoff observed: repeatable success (Observed 20/20) rose sharply from 38 to 107 tasks, while pass@20 (tasks solved at least once) declined slightly from 453 to 446, indicating some tasks became harder to reach at all but far more reliable when solved

Why it matters

The release provides an open, reproducible pipeline for improving agent reliability on structured data tasks, where current models often hallucinate completion. The method is particularly relevant for enterprise agent deployments where database state correctness matters more than tool-call appearance.