Multimodal question answering (MMQA) requires integrating heterogeneous data sources, selecting relevant modalities and ignoring distractors. Existing Multimodal Large Language Model (MLLM) approaches improve performance by decomposing the task into intermediate steps, yet they do not quantify each modality’s contribution. Entropy‑Driven Multimodal Answering (EDMA) formalises MMQA as an iterative entropy‑reduction process using logical entropy, and measures how each modality distinguishes correct from incorrect candidate answers, yielding a quantifiable answering trajectory. The main components are: (i) modality‑conditioned partitioning to estimate unimodal entropy reduction; (ii) entropy‑driven multimodal fusion that captures complementary information across modalities; (iii) entropy‑based selection of the answering trajectory that minimises logical entropy. Evaluated on standard MMQA benchmarks, EDMA outperforms Planning Prompting by 12.8 F1 and 4.6 LLM‑J, and beats Chain‑of‑Thought by 18.6 F1 and 8.3 LLM‑J, surpasses self‑consistency baselines and is competitive with multi‑agent systems. Analysis shows higher entropy correlates with more false positives, and logical entropy serves as a monotonic selector for correct reasoning paths, confirming entropy level as a reliable performance proxy.
Review: EDMA demonstrates that an explicit entropy metric can both improve the interpretability and the accuracy of multimodal reasoning, highlighting the promise of information‑theoretic principles in large‑model multimodal fusion.