theaicatalog
Hugging Face ยท Research

Microsoft and Hugging Face use database checks as rewards to train agents

Microsoft and Hugging Face released ThinkingBox-Training on GitHub. The method uses the final state of a backend database as a reward for reinforcement learning, so an agent is rewarded only when the database ends up correct. The companies say that after 50 GRPO steps on 187 tasks outside their ThinkingBox-Bench benchmark, Qwen3.8-27B completed 107 of 507 benchmark tasks correctly on all 20 attempts, up from 38. Average success rose from 51.70% to 60.78%, while the number of tasks solved at least once fell slightly, from 453 to 446.