A database reward lifts an agent from 38 to 107 tasks done each time

Microsoft and Hugging Face released ThinkingBox-Training on GitHub. It uses a check of the last database condition as a reward to train agents. After 50 steps, the Qwen3.8-27B model went from 51.70% to 60.78% average score on new tasks. Tasks done on all 20 tries went from 38 to 107 of 507.

How it works

ThinkingBox gives an agent a backend and tools. When the agent stops, a check reads the database and gives 1 or 0. The authors used this result as a reward in GRPO, a method that trains a model with a reward. They kept tasks that the first model did right on some tries and not on other tries.

The numbers

GPT-5.6 Sol has an average score of 61.91%, and the tuned model has 60.78%. But GPT-5.6 Sol completes only 82 tasks on all tries against 107. Claude Opus 5.5 is first with 241. Its older version has a lower average score but the same 241.

What is not known

The tuned model did not include more tasks. The number of tasks done on one or more tries fell, and the authors give no number for it. The authors report that the model moved to the procedures that work each time. No other group shows the same result.

Insight: The test author also controls the training reward

Microsoft and Hugging Face used one database check as the score and the reward. Arize gives agent records a score with code checks. MetaExp shows that task records can add skill to a model. The group that writes the test controls the reward, which can move control to test authors. This is our reading. Leaders must find out who controls their task checks.