Fewer tokens, more accuracy: reasoning skills beat reasoning length
A 2026 paper reports that structuring a model's reasoning around explicit skills produces better answers with fewer tokens — cutting against the assumption that more thinking is monotonically better.
"Thinking with Reasoning Skills: Fewer Tokens, More Accuracy" makes a claim worth checking carefully, because it points the opposite way to the last two years of practice: structuring reasoning around explicit skills yields better accuracy while using fewer tokens.
The prevailing assumption has been that reasoning quality scales with reasoning length. It is why models expose thinking budgets, why providers sell reasoning effort as a dial, and why so many production prompts end up begging a model to think step by step. If length is not the operative variable, a lot of spend has been buying tokens rather than accuracy.
The mechanism is intuitive once stated. An unstructured chain wanders — it restates the problem, explores a dead end, recovers. A skill-structured one applies a known procedure. Humans show the same asymmetry: an expert's working is shorter than a novice's, not longer.
The commercial implication is immediate given how inference is now priced. Reasoning tokens are billed, often at a premium, and a technique that cuts them while improving accuracy improves both sides of a ledger that usually trades off.
The caution is that skills have to come from somewhere. A curated skill set is a prior about which procedures matter — powerful where the domain is known, and a ceiling where it is not.
arXiv — Thinking with Reasoning Skills: Fewer Tokens, More Accuracy → · Sebastian Raschka — LLM Research Papers: The 2026 List →