Compose and Fuse: Revisiting the Foundational Bottlenecks in Multimodal Reasoning

ArXi:2509.23744v3 Announce Type: replace Multimodal large language models (MLLMs) promise enhanced reasoning by integrating diverse inputs such as text, vision, and audio. Yet cross-modal reasoning remains underexplored, with conflicting reports on whether added modalities help or harm performance. These inconsistencies stem from a lack of controlled evaluation frameworks and analysis of models' internals to isolate when and why modality interactions or undermine reasoning.