Microsoft releases ThinkingBox-Training for agent RL tuning

Microsoft published ThinkingBox-Training on GitHub, extending its ThinkingBox benchmark by using executable database-state checks as a reward signal for post-training. A Qwen3.8-27B model fine-tuned with GRPO for 50 steps raised pass@1 on the 507-task ThinkingBox-Bench from 51.70% to 60.78%, while pass@20 slipped from 89.35% to 87.97%.

• Pass@1 rose from 51.70% to 60.78%; tasks solved on all 20 runs rose from 7.50% to 21.10%
• Pass@20 fell from 89.35% to 87.97%, so the model got more repeatable, not broader
• Training pool held 187 tasks filtered for mixed outcomes; it is private
• Estimated cost per task passed on every attempt: $8.00 vs $24.36 for the base model