Microsoft published a paper on August 11 introducing a skill-distillation method that enables cheaper AI models to match or outperform their expensive reasoning modes. By analyzing 35 to 50 task trajectories, a coding agent extracts compact rules in markdown documents of 40 to 130 lines, injected into the model’s system prompt. This approach reduces output token usage by 2.7 to 6 times while costing only $1 to $3 per domain, with no model retraining required. On the ALFWorld benchmark, skill-enhanced GPT-5.4-mini achieved a score of 0.787 compared to 0.713 for full reasoning mode, surpassing it. The research team was led by Agamdeep Singh, Srishti Gautam, Priyanshu Gupta, Nikita Mehrotra, Tanmay Bakshi, and Sumit Gulwani.
Source: Read the original article
