Morning Edition · Thursday, July 9, 2026Published at 1:32 AM EDT · New York

New Theory Maps When Extra Test-Time Reasoning Actually Helps LLMs

A sampling-complexity analysis of generate-critique-revise loops formalizes the conditions under which in-context search improves answers rather than only consuming inference compute.

A theoretical paper, When Does In-Context Search Help?, takes on a question that the test-time-compute era has mostly answered by experiment. Extended reasoning, in which a model repeatedly generates, critiques, and revises its own attempts…