Microsoft paper: Distilled skills replace reasoning, cutting tokens by 2.9-4.5x

rohanpaul_ai · x · 2026-08-15

A new Microsoft paper proposes reducing test-time reasoning costs by extracting recurring failure patterns from past agent trajectories and turning them into small markdown skills added to the system prompt of a non-reasoning model. Experiments show this recovers 55%-100%+ of the gap between non-reasoning and reasoning modes across 4 agent benchmarks, while using 2.9-4.5x fewer output tokens. On ALFWorld and τ²-retail, the skilled non-reasoning model even beat the reasoning mode. The distiller did not need reasoning traces; skills built only from cheap non-reasoning rollouts were competitive across all 4 domains. However, reasoning still won on telecom and SpreadsheetBench, where tasks have more instance-specific dependencies. Practical takeaway: distill repeated procedures once, reserve expensive reasoning for tasks needing fresh search.

Original post →

More from coding & agent

coding & agent channel →