ThinkingBox turns database grading into RL rewards for post-training agents

Researchers from Microsoft and Hugging Face introduced ThinkingBox, a system that evaluates agent performance by comparing final database states with expected outcomes. Using executable checks as reward signals, they fine-tuned Qwen3.8-27B with GRPO for 50 steps. This training significantly improved repeatable success, increasing pass@1 from 51.70% to 60.78% and tasks completed correctly on all 20 attempts from 38 to 107.