On October 22, Microsoft and Hugging Face released ThinkingBox Training for database-state rewards.

The big change

  • What changed: Database-state checks became agent training rewards.
  • Why it matters: The authors report 107 of 507 tasks passed all 20 attempts, up from 38. Engineers can weigh consistency.
  • What to watch: At-least-once coverage fell from 453 to 446. Check breadth; live deployment performance remains unmeasured.

After 50 GRPO steps with Qwen3.8-27B, the authors counted 6,163 successes in 10,140 attempts, up from 5,242.

Base-model probes fixed a private 187-task pool; the public 39-task sample differs. Most rewards used executable checks, some a judge.

Reference run targets Linux, Python 3.12, patched VeRL 0.9.0 and 24 H100 GPUs, with simulated-user and judge endpoints. Training cost is unreported.