RESEARCH · RESEARCH · #2316
EDMA: entropy-driven framework for multimodal question answering
EDMA (Entropy-Driven Multimodal Answering) is an arXiv v1 paper that formalises multimodal question answering as an iterative process of logical-entropy reduction. It proposes modality-conditioned partitioning, entropy-driven multimodal fusion, and entropy-based trajectory selection to quantify each modality's contribution and selects low-entropy reasoning paths; evaluated on standard MMQA benchmarks, EDMA reports average gains of 12.8 F1 and 4.6 LLM-J over Planning Prompting and 18.6 F1 and 8.3 LLM-J over Chain-of-Thought, and shows higher entropy correlates with more false positives.
KEY POINTS
- EDMA (Entropy-Driven Multimodal Answering) is an arXiv v1 paper that formalises multimodal question answering as an iterative process of logical-entropy reduction.
- It proposes modality-conditioned partitioning, entropy-driven multimodal fusion, and entropy-based trajectory selection to quantify each modality's contribution and selects low-entropy reasoning paths; evaluated on standard MMQA benchmarks, EDMA reports average gains of 12.8 F1 and 4.6 LLM-J over Planning Prompting and 18.6 F1 and 8.3 LLM-J over Chain-of-Thought, and shows higher entropy correlates with more false positives.
- Provides a measurable entropy-based method to attribute modality contributions and select reasoning trajectories, improving traceability and performance in multimodal QA.
WHY IT MATTERS
Provides a measurable entropy-based method to attribute modality contributions and select reasoning trajectories, improving traceability and performance in multimodal QA.