Microsoft and Hugging Face report ThinkingBox training boosts Qwen repeatability but narrows task coverage

In an October 22 report, Microsoft and Hugging Face say that ThinkingBox uses final database-state checks as reinforcement-learning rewards. After 50 GRPO steps, Qwen3.8-27B average success increased from 51.70% to 60.78%, and tasks passed on all 20 attempts rose from 38 to 107 (source).

The tradeoff matters for database-writing agents: Microsoft and Hugging Face report that tasks solved at least once fell from 453 to 446, supporting their recommendation to assess repeatability alongside coverage rather than average success alone (source).

Microsoft and Hugging Face identify limitations: training and evaluation share infrastructure and task style, limiting transfer claims; benchmark tasks are synthetic, and public samples exclude the private 187-task training pool. These vendor-reported results leave production reliability and reproducibility from public resources unestablished (source).