将数据库评分转化为 RL 奖励：使用 ThinkingBox 对智能体进行后训练

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

We used the executable terminal-state checks from ThinkingBox-Bench as rewards for RL fine-tuning, selecting tasks based on offline policy-relative difficulty to improve agent dependability without sacrificing coverage.