Anatomy of a Hybrid Mind: Deconstructing Hybrid Reasoning in Large Language Models

ICLR 2026 Conference Submission15261 Authors

Keywords: Hybrid Reasoning

TL;DR: We conduct a mechanistic analysis to uncover how hybrid reasoning modes coexist and switch in LLMs.

Abstract:

Hybrid reasoning, which enables a *single* Large Language Model (LLM) to alternate between fast, intuitive responses (non-thinking mode) and slow, deliberate reasoning (thinking mode), has rapidly gained adoption across the AI industry, spanning from top-tier commercial models to the latest open-source releases. Despite their widespread deployment, the community still lacks a mechanistic understanding of how these two modes coexist and interact *within a single model*. Without such clarity, we lack a principled understanding of how these modes operate, making it harder to reason about their behavior or guide future development. In this work, we conduct a detailed mechanistic analysis of a hybrid reasoning model’s internal dynamics. We identify why the thinking and non-thinking modes can be compatible rather than distinct subsystems. Building on this, we propose a metacognitive taxonomy showing that the thinking mode corresponds to a structured, self-corrective protocol whose intensity can be modulated along a continuous spectrum. Furthermore, we causally uncover a surprisingly localized, single-token switch that deterministically governs mode activation.
These findings illuminate the control mechanisms underlying hybrid reasoning, providing a foundation for the design of robust, interpretable, and adaptive cognitive systems.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 15261