Microsoft Cracks Code on Cheaper AI Models Outperforming Reasoning Mode

Microsoft researchers have made a breakthrough in artificial intelligence by finding a way to make a cheaper AI model outperform its expensive reasoning mode. The team used a passive skill-distillation method, which extracts compact rule sets from a small number of task examples and injects them into the model's system prompt.

The process is deceptively simple, involving a coding agent reviewing between 35 and 50 task trajectories to distill a natural-language 'skill' document ranging from 40 to 130 lines of markdown. These rules are derived from what went wrong and what went right in previous attempts and can be audited.

The compute cost for generating each domain's skill document runs between $1 and $3, with no model retraining required. The team evaluated the approach across four agentic benchmarks, including ALFWorld and SpreadsheetBench-Verified, and found that the cheaper model used 2.7 to 6 times fewer output tokens than the reasoning mode.

The new distillation method builds directly on Microsoft's SkillOpt framework, which reframed agent skills as trainable parameters rather than static model weights. This concept has been further made accessible with the new skill-augmented model working seamlessly with existing rollout strategies and no custom inference infrastructure required.