Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox

Microsoft and Hugging Face are using ThinkingBox's executable database checks as reinforcement-learning rewards to post-train agents, and report that Qwen3.8-27B's average success rose from 51.70% to 60.78% after 50 GRPO steps.

• ThinkingBox was built to catch agents that claim they are done when the database says otherwise, and its 507-task benchmark hands an agent a real backend, tools and a simulated user before checking the terminal database state.
• The authors say most tasks are either already solved or hopeless for a given model, so they estimated each task's success probability once with the base policy and kept that training pool fixed, an approach they call offline policy-relative difficulty.
• On the sealed benchmark, tasks completed correctly on all twenty attempts rose from 38 to 107, while at-least-once coverage declined slightly.
• In one hotel-booking task, the base model changed a booking without required manager approval in 15 of 20 attempts, while the fine-tuned model preserved booking and payment state in all 20.