Amortizing Reasoning Premium: Skills Distilled from Trajectories

A recent submission on arXiv (2608.07885) introduces a technique aimed at reducing the computational burden associated with reasoning in language models by extracting reusable skills from prior trajectories. The researchers note that while reasoning modes in language models excel in multi-step agentic tasks, they require 3-6 times more output tokens per episode, primarily due to the need to re-derive shared procedures across similar domain episodes. To mitigate this issue, a coding agent evaluates a limited set of existing trajectories from a training split and formulates a concise natural-language skill to enhance the non-reasoning model's system prompt. In tests across four agentic benchmarks (ALFWorld, tau^2-bench telecom and retail, and SpreadsheetBench-Verified), these skills bridge 55%-100%+ of the reasoning gap for GPT-5.4-mini on held-out tasks, surpassing the reasoning mode in two out of four benchmarks while producing 2.7-6 times fewer output tokens and no reasoning tokens. Interestingly, skills derived solely from non-reasoning trajectories remain effective, even without reasoning traces.