When Does In-Context Search Help? A Sampling-Complexity Theory of Reflection-Driven Reasoning

A new theoretical analysis of in-context search in large language models shows that when self-reflection reliably localizes early mistakes, sequential reasoning can yield exponential improvements over the base model, solving problems with exponentially small zero-shot pass rates using only a polynomial number of attempts. The study further demonstrates that these gains are robust and learnable, with cross-entropy training on search rollouts recovering the required behavior with polynomial sample complexity.

arXiv:2607.06720v1 Announce Type: new Abstract: Training large language models LLMs with extended reasoning has enabled in-context search, in which models iteratively generate, critique, and revise solution attempts. We provide a theoretical analysis of in-context search by modeling it as approximate inference over reasoning traces, where the base model defines a prior and self-reflection provides feedback for posterior updates, and study the resulting inference-time sampling complexity - the number of sequential attempts needed to achieve high success probability. We show that when reflections reliably localize early mistakes, in-context search can yield exponential improvements over the base model, solving problems with exponentially small zero-shot pass rates using only a polynomial number of sequential attempts, whereas when this property fails, conditioning on past attempts offers no asymptotic benefit over parallel sampling. We further show that these gains are robust and learnable: approximate posterior updates suffice, and cross-entropy training on search rollouts recovers the required behavior with polynomial sample complexity. Finally, we show that under a stagewise abstraction of reinforcement learning with verifiable rewards, the optimal policy extension implements the same posterior reweighting rule. We validate key qualitative predictions of the theory on real large reasoning models.