ThinkingBox Turns Database-State Checks into Agent Training Rewards
- Microsoft and Hugging Face have released ThinkingBox-Training on GitHub, including the data-selection, rollout, reward, and Verl integration used in their Qwen3.8-27B experiment.
- Instead of grading a sequence of tool calls, its reward loop checks whether the backend ends in the required state; most checks are deterministic, while a smaller portion use an LLM judge.
- After 50 GRPO steps, Qwen3.8-27B’s pass@1 average success rate rose from 51.70% to 60.78%.
- The number of benchmark tasks passed on every attempt increased from 38 to 107, while the share passed at least once declined slightly.
- Training used a fixed pool of 187 tasks chosen through offline probes of the base model, and evaluation covered 507 ThinkingBox-Bench tasks excluded from training, with 20 trials per task.
Our take
This is not a drop-in replacement for generic RL: teams need task-specific executable state checks, an isolated backend, and aligned tool parsing in training and evaluation. The reported serving-cost comparison excludes training and task-probing costs, so it does not establish the end-to-end economics of this workflow against other post-training approaches.
Sources
- Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox · Hugging Face Blog
The summary and analysis were generated by AI. The original sources are authoritative. AI content policy