Definition of a Meaningful Interleaved Chain-of-Thought

Determine rigorous and operational criteria that define a meaningful interleaved chain-of-thought in multimodal reasoning, specifying how textual tokens and image tokens should interact as complementary modalities to mutually advance reasoning beyond mere isomorphic representations, and establishing evaluation procedures to verify these criteria across diverse tasks.

Background

The paper motivates interleaved multimodal reasoning by noting that current models struggle when problems demand more than textual description. While textual Chain-of-Thought has advanced verbal reasoning, the authors argue it contributes little to multimodal reasoning without concrete visual manipulation steps that complement text. They therefore emphasize that what counts as meaningful interleaving—where text and image thoughts truly reinforce each other—has not been clearly defined.

To address the gap empirically, the authors construct ThinkMorph and curate ~24K high-quality interleaved traces across tasks requiring varying levels of visual engagement. Despite these efforts, the abstract explicitly highlights the unresolved need for a principled definition and criteria for meaningful interleaving, underscoring the broader methodological open problem.