A multimodal cross-modal explainable retrieval-augmented generation framework for hallucination-grounded visual question answering

Abstract

Recent multimodal large language models (MLLMs) have achieved remarkable performance on visual question answering (VQA) and multimodal reasoning tasks. But one overlooked failure case stubbornly persists: when answers require simultaneous synthesis of evidence from natural images, free-form text, and structured tables, present models suffer from trimodal hallucination—generating content unattributed to any of the input modalities. Prior approaches to hallucination mitigation focus on image-text pairs, and existing retrieval-augmented generation (RAG) systems for multimodal settings do not readily include supporting evidence and often lack modality-attributed explainability. In this paper, we present MXRAG (Multimodal Cross-Modal Explainable Retrieval-Augmented Generation), a new approach to this trimodal evidence problem with three key innovations that work in concert: (1) a Trimodal Evidence Retriever (TER) that retrieves image patches, text passages, and table rows jointly using a shared semantic manifold; (2) a Cross-Modal Attribution Network (CMAN) that computes fine-grained, token-level attribution scores inline during generation, mapping each generated token to supporting evidence from all three modalities; and (3) a Hallucination-Aware Constrained Decoding (HACD) strategy that penalises generation steps with attribution entropy above a calibrated threshold, suppressing unsupported factual tokens at inference time. CMAN is trained with novel cross-modal attribution and modality-coherence losses using token-level gold annotations; HACD requires no additional training and is calibrated per dataset on the validation split. We cast joint retrieval-generation as a constrained variational problem over a trimodal evidence space and introduce MMTabQA, a new trimodal VQA benchmark derived from WikiTableQuestions and MSCOCO with 12,847 instances and token-level attribution labels. Evaluation on four benchmarks (MMTabQA, WebSRC, ChartQA, MIMIC-CXR-VQA) shows that MXRAG achieves state-of-the-art. +21.4% points (p.p.) exact match accuracy, − 38.6% relative reduction in hallucination rate (− 12.1 p.p. absolute) versus the best multimodal RAG baseline, and 89.4% modality coherence score. Ablation experiments confirm the contribution of each component, with CMAN providing the largest accuracy gain (+ 14.2%) and HACD the largest hallucination reduction (− 23.1%). A supplementary human evaluation on 200 instances confirms that the entropy-based hallucination metric tracks human judgement (human-judged HR: 21.3% vs. metric HR: 19.1%). MXRAG advances the state of the art for reliable, interpretable, evidence-based multimodal AI.