Understand this story
The research team performed full-parameter RL fine-tuning of Qwen3.8-27B, using the executable checks on the final database state in ThinkingBox-Bench directly as reinforcement learning rewards. This work was done by the Microsoft Copilot Studio team in collaboration with Toloka and others. On 507 out-of-distribution tasks not used in training, the number of tasks the model passed on all 20 attempts rose from 38 (7.5%) to 107 (21.1%), pass@1 rose from 51.70% to 60.78%, and the cost per successful attempt dropped from $0.177 to $0.139. The training pool used offline-policy relative-difficulty filtering, ultimately retaining 187 tasks; the training and evaluation code has been open-sourced at microsoft/thinkingbox-training.
AI-generated from public reports
Latest development · Report
Microsoft uses ThinkingBox to turn database scoring into RL rewards, fine-tuning Qwen3.8-27BDevelopments
1 confirmed developments
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 23
Hugging FaceFirst-party
Microsoft uses ThinkingBox to turn database scoring into RL rewards, fine-tuning Qwen3.8-27B
The Microsoft Copilot Studio team, in collaboration with Toloka and others, directly uses the executable final-state database checks of ThinkingBox-Bench as reinforcement learning rewards to perform full-parameter RL fine-tuning on Qwen3.8-27B. On 507 out-of-distribution tasks not involved in training, the number of tasks the model passed in all 20 attempts rose from 38 (7.5%) to 107 (21.1%), pass@1 rose from 51.70% to 60.78%, and the cost per successful attempt dropped from $0.177 to $0.139. The training pool was filtered using offline-policy relative difficulty, ultimately retaining 187 tasks; the training and evaluation code has been open-sourced at microsoft/thinkingbox-training.