Microsoft and Hugging Face boost AI agent consistency with database-based training

Microsoft and Hugging Face have turned a check on AI agents’ work into a training reward: did the database end up correct? In an October 22 joint post, they reported that this approach made Qwen3.8-27B more consistent on business tasks, while slightly reducing the number it could solve at least once.

The new ThinkingBox-Training release uses executable checks of final database records to guide reinforcement learning—training through rewards. The team ran 50 steps on 187 tasks outside ThinkingBox-Bench, then evaluated the resulting model on the benchmark’s 507 tasks, with 20 attempts per task.

Reward the outcome, not a prescribed route

ThinkingBox checks what an agent leaves behind rather than requiring a particular sequence of tool calls. Its underlying research paper describes checks that reject wrong, missing or unwanted changes while accepting valid routes to the result. Some tasks also test required properties of the final response.

For training, most rewards came from deterministic database checks: code returned a success or failure verdict. A minority of narrowly defined requirements could invoke GPT-5 Chat as a judge. Each attempt received a fresh backend and isolated tool session, with GPT-5 Chat also acting as a simulated user.

Task selection supplied the learning contrast. The team estimated difficulty using the base model before training and kept that pool fixed. Its GRPO training method compares outcomes within groups: eight failures or eight successes offer no difference to learn from, while mixed results do. The reference run used 24 H100 GPUs.

More repeatable is not the same as broader

The authors describe the pattern as consistent with sharpening: successful behavior becomes more repeatable while some less-likely routes to success disappear. That is an explanation, not a demonstrated cause. Average success and all-20 success improved across all five benchmark domains; at-least-once coverage fell in retail, travel and auto insurance.

The distinction also changes model comparisons. In the team’s table, GPT-5.6 Sol had a higher average success rate than the trained Qwen model, 61.91% versus 60.78%. Yet it passed fewer tasks on every recorded attempt: 82 versus 107. Neither average accuracy nor repeated success alone describes the full result.

One hotel-booking task makes the behavioral change concrete. A requested room downgrade required manager approval. Before training, the model changed the booking and issued refunds without approval in 15 of 20 attempts. Afterward, it preserved booking and payment state every time—but matched all expected records in 19 attempts, not 20.

Lower estimated serving cost, with boundaries

Using measured token usage, the team estimated that evaluation cost per task passed on all 20 attempts fell from $24.36 to $8.00. These figures are a comparative serving index, not cloud bills. They exclude training and the preliminary model runs used to select tasks, so they do not measure total project cost.

The held-out benchmark used domains, databases, policies and most tools absent from training. But both sets shared the ThinkingBox framework, simulated-user machinery, grading approach, some tools and a similar task style. The authors explicitly limit their transfer claim on that basis. ThinkingBox-Training is available on GitHub, including task selection, agent-run generation, rewards and training integration.