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

What happened

Microsoft and Hugging Face released ThinkingBox-Training on GitHub, turning executable checks of final database states into rewards for agent post-training. After 50 GRPO steps on 187 tasks outside ThinkingBox-Bench, Qwen3.8-27B’s average benchmark success rose from 51.70% to 60.78%, and tasks passed on all 20 attempts increased from 38 to 107, while tasks solved at least once fell from 453 to 446. The release includes task selection, rollout, reward, and Verl integration code; the reported run’s private training pool is not included.

Read next

• More changes in today’s briefingOlmoEarth v1.3 releases global satellite embeddings and a browser-based mapping explorer

The story so far

Microsoft and Hugging Face released ThinkingBox-Training on GitHub, a workflow that turns executable checks of final database states into rewards for agent post-training. Hugging Face reported that after 50 GRPO steps on 187 tasks outside ThinkingBox-Bench, Qwen3.8-27B’s average benchmark success rose from 51.70% to 60.78%. Tasks passed on all 20 attempts increased from 38 to 107, although tasks solved at least once fell from 453 to 446. The release includes task-selection, rollout, reward, and Verl integration code. Its public sample task list exercises the interface; the private 187-task training pool used for the reported results is not included. Running the workflow requires NVIDIA GPUs with a compatible CUDA/NCCL stack, a local Qwen3.8-27B checkpoint, and OpenAI-compatible endpoints for the simulated user and judge. The validated reference configuration uses three nodes with eight H100 GPUs each.