Evidence Selection and Structured Reasoning forKnowledge-Intensive Multimodal Question Answering: A Survey

The Hong Kong University of Science and Technology (Guangzhou)

Data Science and Analytics Thrust

PhD Qualifying Examination

By Mr. LIU, Bingcen

Abstract

Knowledge-intensive multimodal question answering (KI-MQA) requires systems to connect visual observations with external knowledge and multi-step inference. This survey argues that reliable KI-MQA depends on jointly designing evidence selection and structured reasoning. Relevant context may be incomplete or useless, while fluent reasoning traces may remain unsupported or post hoc. We organize evidence-selection methods by source, operation, and decision signal, and reasoning methods by representation, control strategy, and grounding strength. The literature shows a shift from fixed retrieval and static reasoning toward question-conditioned visual compression, answer-aware retrieval, reflective filtering, and adaptive tool use. Yet retrieval recall, visible rationales, and source attribution remain incomplete proxies for reliability. Emerging research directions point toward an evidence-state perspective that explicitly models selected evidence, claim dependencies, provenance, unresolved needs, and source reliability to enable more reliable multimodal reasoning.

PQE Committee

Chair: Prof. YU, Xu Jeffrey

Prime Supervisor: Prof. CHEN, Lei

Co-Supervisor: Prof. ZHANG, Yongqi

Examiner: Prof. ZHU, Lei