![[2509.23744] Compose and Fuse: Revisiting the Foundational Bottlenecks in Multimodal Reasoning](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2509.23744] Compose and Fuse: Revisiting the Foundational Bottlenecks in Multimodal Reasoning
Summary
This article explores the foundational bottlenecks in multimodal reasoning, highlighting how additional modalities can enhance or hinder performance in multimodal large language models (MLLMs).
Why It Matters
Understanding the complexities of multimodal reasoning is crucial for advancing AI capabilities. This research identifies key failures in current models and suggests new training approaches, which could lead to more effective integration of diverse data types in AI systems.
Key Takeaways
- Multimodal reasoning can improve performance if additional modalities provide independent reasoning paths.
- Redundant or chained entailments often degrade reasoning quality.
- Two core bottlenecks identified: task-composition and fusion bottlenecks.
- Attention patterns currently fail to encode the usefulness of facts.
- Composition-aware training and improved early fusion techniques are recommended.
Computer Science > Computation and Language arXiv:2509.23744 (cs) [Submitted on 28 Sep 2025 (v1), last revised 25 Feb 2026 (this version, v2)] Title:Compose and Fuse: Revisiting the Foundational Bottlenecks in Multimodal Reasoning Authors:Yucheng Wang, Yifan Hou, Aydin Javadov, Mubashara Akhtar, Mrinmaya Sachan View a PDF of the paper titled Compose and Fuse: Revisiting the Foundational Bottlenecks in Multimodal Reasoning, by Yucheng Wang and 4 other authors View PDF Abstract: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 support or undermine reasoning. We address this gap through a logic-grounded evaluation framework that categorizes multimodal reasoning into six interaction patterns, varying how facts are distributed across modalities and logically combined. Empirically, additional modalities enhance reasoning only when they provide independent and sufficient reasoning paths, while redundant or chained entailment support often hurts performance. Moreover, reasoning degrades in three systematic ways: weaker modalities drag down overall performance, conflicts bias preference toward certain modalities, and joint s...
