ThinkingBox rewards agents for the database state they leave behind
Executable outcome checks become training rewards. The reported gains in repeatability come with limits on coverage and transfer.
An agent hands back a completed task. The database may tell a different story. ThinkingBox turns that gap into a training signal by checking the final state and using the outcome as a reward.
In the article published October 22, 2026, Tuhin Kundu and collaborators describe deterministic checks for most tasks. Incorrect changes, missing effects, and unwanted extra effects fail those checks. Narrow requirements that database state cannot adequately represent may use a language-model judge. Each attempt runs in an isolated environment.
The reported training improved repeatability on the evaluation set while slightly reducing the range of tasks solved at least once. Training and evaluation tasks were separate, but they shared infrastructure and task style. That connection limits claims about transfer to unrelated operating environments.
I would define completion alongside the intended system change, then record the state that demonstrates the handoff succeeded. The remaining operational choice is which requirements can become executable checks and which still need judgment.
Source: Tuhin Kundu and the Microsoft and Hugging Face team, the ThinkingBox-Training article. The project is developed in partnership with Toloka.