Artificial intelligence / news brief
Microsoft and Hugging Face turn database checks into agent-training rewards
- Event
- Published
Microsoft and Hugging Face released ThinkingBox-Training, which uses executable checks of an agent's final database state as reinforcement-learning rewards. In a 507-task evaluation, the authors report Qwen3.8-27B's average success rising from 51.70% to 60.78%, while tasks solved at least once fell from 453 to 446.
Why it matters
Outcome-based rewards can train enterprise agents against the records they actually change, while the coverage drop shows that repeatability gains do not improve every task.
Sources
- Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox (external site)Microsoft and Hugging Face · primary source ·
- Microsoft trains Qwen3.8-27B on 187 tasks to improve agent repeatability (external site)Runtime Wire · coverage ·