Reason Wide, Not Deep: Amortizing the Reasoning Premium into Distilled Skills

2026-08-07

AI Summary

This work addresses the high token cost incurred by language models in multi-step agent tasks, where repeated deep reasoning leads to output token overheads of 3–6×. The authors propose a novel “broad search” paradigm that distills compact natural-language skills from a small set of training trajectories and injects them into the system prompt, enabling procedural knowledge reuse across turns without per-turn re-reasoning. This approach demonstrates, for the first time, that effective skills can be extracted without access to explicit reasoning trajectories. Evaluated on four agent benchmarks, it recovers 55%–100%+ of the original reasoning performance—surpassing native reasoning in two tasks—while reducing output tokens by 2.7–6× and entirely eliminating per-turn reasoning overhead.

Abstract

Reasoning modes of language models outperform their non-reasoning counterparts on multi-step agentic tasks, but pay a 3-6x premium in output tokens on every episode -- much of it spent re-deriving procedures that are shared across episodes of the same domain. We show this recurring cost can be amortized: a coding agent analyses a small corpus of existing trajectories from a training split and compiles a compact natural-language skill that is injected into the non-reasoning model's system prompt. Across four agentic benchmarks (ALFWorld, tau$^2$-bench telecom and retail, and SpreadsheetBench-Verified), skills recover 55%-100%+ of the reasoning gap for GPT-5.4-mini on held-out tasks -- exceeding the reasoning mode outright on two of four -- while emitting 2.7-6x fewer output tokens and zero reasoning tokens. Notably, reasoning traces are not a prerequisite: skills distilled from non-reasoning trajectories alone remain competitive with skills distilled from paired reasoning/non-reasoning corpora, with domain-dependent differences between the two sources. We interpret these results through a search lens: test-time reasoning is deep search inside a single episode, re-paid at every deployment, while corpus distillation is wide search across episodes, paid once. The two recover overlapping procedural knowledge, and width over cheap trajectories is often the better buy -- with the residual gap on some domains (telecom, SpreadsheetBench) delineating where genuinely per-instance deep search remains necessary.

Problem

• reasoning cost
• token efficiency
• procedural knowledge
• agentic tasks
• amortization

Innovation

• amortized reasoning
• skill distillation
• agentic benchmarks
• natural-language skills
• width vs depth