Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox

Microsoft introduces ThinkingBox-Training, a post-training method for AI agents that uses executable checks of a database's final state as reinforcement learning rewards. According to the announcement, this approach improved Qwen3.8-27B's performance on tasks outside the ThinkingBox-Bench benchmark: average success rate rose from 51.70% to 60.78%, and the number of tasks completed correctly on all 20 attempts increased from 38 to 107 after 50 GRPO steps.

ThinkingBox-Bench evaluates agents by running a real backend and verifying the final database state, regardless of the tool calls made. The benchmark's 507 tasks simulate real-world scenarios like booking management or customer ticket handling. The team used the executable outcomes of these tasks to build a tailored training pool, selecting tasks whose difficulty is relative to the model's current policy. The observed gains show improved repeatable reliability, with a slight decline in overall coverage.

The results highlight that repeatability (tasks solved on every attempt) is a key production criterion, distinct from average performance. The fine-tuned model reduces the estimated cost per reliable task to $8, approaching larger models like GPT-5.4 or Claude Opus 5.5, while remaining cheaper to serve.