Microsoft and Hugging Face Release ThinkingBox-Training to Turn Database State Checks into Agent Rewards

On October 23, 2026, Microsoft and Hugging Face released ThinkingBox-Training, an open-source training stack available on GitHub. It converts executable checks of terminal database state into reinforcement learning rewards for post-training agents. The work is a joint effort between the two organizations.

ThinkingBox was originally built to catch agents that report task completion without making the intended changes to a database. In the new project, the same kind of executable signal is used as a reward for training tasks outside ThinkingBox-Bench.

To choose training tasks, the team applied an offline policy-relative difficulty filter: each candidate task was probed repeatedly with the base policy, and tasks with success probabilities likely to produce contrasting reward groups were retained. The final pool contained 187 tasks for Qwen3.8-27B. The model was fine-tuned with GRPO for 50 full-parameter steps.

After training, average success (pass@1) on the 507-task ThinkingBox-Bench rose from 51.70% to 60.78%, and tasks solved on all 20 attempts rose from 38 to 107. Tasks solved at least once fell slightly from 89.35% to 87.97%. The release includes documentation for data preparation, evaluation, and the validated 24-H100 training configuration.