EDMA: Entropy-Driven Multimodal Answering, a new preprint on arXiv, tackles a core challenge in multimodal question answering (MMQA): pulling together heterogeneous data while ignoring irrelevant or distracting information. The abstract frames MMQA as a problem of selective modality use, where models must decide which sources matter for a given question.
Existing approaches built on multimodal large language models (MLLMs) are the baseline, but the abstract suggests they fall short in this selective integration. The name EDMA points to an entropy-driven mechanism, likely using uncertainty or information content to guide which modalities to trust. The full technical details are not visible in the listing, but the framing indicates a principled alternative to current MLLM-based methods.
The paper is available on arXiv under identifier 2610.27120.