Microsoft and Hugging Face released ThinkingBox-Training, an open-source codebase for using terminal database states as binary reward functions during agent post-training. By filtering tasks using an offline policy-relative difficulty measure and applying 50 steps of GRPO, they fine-tuned Qwen3.8-27B to increase average success from 51.7% to 60.8%. Tasks completed successfully across all twenty test attempts rose from 38 to 107.
It is worth adopting if you are already running GRPO loops and need objective, non-LLM-as-a-judge verification for backend tool use. The catch is that you need a reliable execution sandbox for your target databases and enough compute to filter out tasks that are either too easy or hopeless for your base model. Skip this if your agent workflows do not involve persistent state changes or complex multi-step backend transactions.
more: https://huggingface.co/blog/microsoft/thinkingbox-training
Leave a Reply