Microsoft open-sources ThinkingBox, lifting Qwen3.8-27B pass@1 from 51.7% to 60.8%

Microsoft released ThinkingBox, a framework and 507-task benchmark for training agents with database-graded rewards, plus training code (MIT) and data (CDLA-Permissive-2.0). Reinforcement fine-tuning Qwen3.8-27B took pass@1 from 51.70% to 60.78% and cost per dependable task from $24.36 to $8.00. Results are Microsoft's own; the 187-task training pool is not included.

HEAT

OUTLETS

0

FIRST SEEN

Oct 22, 2026, 18:24 ET

LAST SIGNAL

3d ago