Training-Free Retrieval-Augmented Generation for Knowledge-Intensive Visual Question Answering

Shan Ning, Longtian Qiu, Jiaxuan Sun, Xuming He

Keywords: Retrieval-Augmented Generation, Visual Question Answering, Multi-modal Large Language Model

Abstract:

Recent advancements in multimodal large language models (MLLMs) have achieved strong performance in vision-language tasks such as visual question answering (VQA). However, these models struggle with knowledge-intensive VQA (KI-VQA) tasks that require fine-grained domain knowledge, as seen in benchmarks such as Encyclopedic VQA and InfoSeek. To address these challenges, we propose a novel retrieval-augmented generation (RAG) framework, referred to as KIRA, designed to enhance the capability of MLLMs for KI-VQA without task-specific fine-tuning. Our target is to integrate general image-text similarity with detailed knowledge context to achieve precise entity recognition. To this end, we leverage CLIP to obtain general image-text matching, and design a verification mechanism according to detailed question-text relevance to improve recognition accuracy. We evaluate our method on KI-VQA benchmarks, demonstrating significant improvements of 47.5% on Encyclopedic VQA and 16.2% on InfoSeek, all achieved without additional training. These results highlight the potential of our training-free, plug-and-play framework for solving knowledge-intensive visual question answering tasks.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6592