Microsoft paper: distilled trajectory skills recover 55-100% of reasoning gains at 2.7-6x fewer tokens
rohanpaul_ai · x · 2026-09-07
- Microsoft researchers propose amortizing expensive test-time reasoning into reusable natural-language "skills" (arXiv 2608.07885)
- Reasoning modes beat non-reasoning on multi-step agentic tasks but pay a 3-6x output-token premium each episode, much of it re-deriving shared procedures
- Method: collect 35-50 past trajectories, have a coding agent extract recurring failure patterns, compile them into a compact skill injected into the non-reasoning model's system prompt
- Across ALFWorld, tau²-bench (telecom/retail) and SpreadsheetBench-Verified, skills recover 55%-100%+ of the reasoning gap for GPT-5.4-mini on held-out tasks, exceeding reasoning mode outright on two of four, with 2.7-6x fewer output tokens and zero reasoning tokens
- Surprisingly, skills distilled from non-reasoning trajectories alone stay competitive with those from paired reasoning corpora
- Framed as search: test-time reasoning is deep search re-paid per deployment; corpus distillation is wide search paid once
Related event: Microsoft Paper: Distilled Skill Cards Match Test-Time Reasoning(2 posts)→
More from coding & agent
- Indie dev's agent tracks 519 directories, 311 submitted for SEO backlinks — fekdaoui · 2026-10-23
- ScienceClaw Hackathon at MIT Media Lab to build decentralized AI agent collectives for science — ProfBuehlerMIT · 2026-10-23
- What should a native agent harness handle beyond tool calling? — Glittering812 · 2026-10-23
- Running ~100 DeepSeek v4.1 flash agents in parallel for under $20/month — zack_overflow · 2026-10-23
- MCP's stateless 2026-07-28 spec drops session-id — how are stateful tools adapting? — Stock-Pumpkin-8859 · 2026-10-23
- Dev adds local MCP server to Mac screenshot app so Claude designs from your references — Embarrassed-Ebb-740 · 2026-10-23