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

Executable checks of a final database state can be used as a reward function for reinforcement learning to train AI agents. This method was implemented using the ThinkingBox framework to post-train a Qwen3.8-27B model on tasks outside of the main benchmark. After 50 training steps, the model's average success rate on the benchmark increased from 51.70% to 60.78%. The training significantly improved reliability, nearly tripling the number of tasks completed correctly on all twenty attempts from 38 to 107. This approach used a concept called "offline policy-relative difficulty" to select suitable training tasks by estimating their success probability before training.