Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox
huggingface ·
Microsoft and Hugging Face have released ThinkingBox-Training, a framework that converts database state verification into reinforcement learning rewards for AI agents. By using executable checks of terminal database states as binary rewards, the system trains models to ensure they produce correct final outcomes rather than just mimicking tool calls. The authors applied this method to fine-tune Qwen3.8-27B using Group Relative Policy Optimization (GRPO) over 50 steps. This training significantly improved the model's reliability, increasing the percentage of tasks completed successfully on all twenty attempts from 7.5% to 21.1%. While average success rates rose by nearly nine points, there was a slight decline in tasks solved at least once, indicating a shift toward more consistent performance. The team utilized offline policy-relative difficulty estimation to select a fixed pool of 187 challenging tasks for training, excluding the benchmark data entirely. Results show the fine-tuned model achieved lower costs per dependable task compared to its base version, approaching the efficiency of larger proprietary models.
huggingface