{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/focus-internal-mllm-representations-for","title":"FOCUS: Internal MLLM Representations for Efficient Fine-Grained Visual Question Answering","arxiv_id":null,"date":"2025-06-25","proceeding":null,"authors":["Liangyu Zhong","Fabio Rosenthal","Joachim Sicking","Fabian Hüger","Thorsten Bagdonat","Hanno Gottschalk","Leo Schwinn"],"abstract":"While Multimodal Large Language Models (MLLMs) offer strong perception and\r\nreasoning capabilities for image-text input, Visual Question Answering (VQA)\r\nfocusing on small image details still remains a challenge. Although visual cropping\r\ntechniques seem promising, recent approaches have several limitations: the need\r\nfor task-specific fine-tuning, low efficiency due to uninformed exhaustive search,\r\nor incompatibility with efficient attention implementations. We address these\r\nshortcomings by proposing a training-free visual cropping method, dubbed FOCUS,\r\nthat leverages MLLM-internal representations to guide the search for the most\r\nrelevant image region. This is accomplished in four steps: first, we identify the\r\ntarget object(s) in the VQA prompt; second, we compute an object relevance map\r\nusing the key-value (KV) cache; third, we propose and rank relevant image regions\r\nbased on the map; and finally, we perform the fine-grained VQA task using the top\u0002ranked region. As a result of this informed search strategy, FOCUS achieves strong\r\nperformance across four fine-grained VQA datasets and two types of MLLMs. It\r\noutperforms three popular visual cropping methods in both accuracy and efficiency,\r\nand matches the best-performing baseline, ZoomEye, while requiring 3 – 6.5×\r\nless compute.","url_abs":"https://arxiv.org/abs/2506.21710","url_pdf":"https://arxiv.org/pdf/2506.21710","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[],"tasks":[{"task_slug":"question-answering","task_name":"Question Answering"},{"task_slug":"visual-question-answering-1","task_name":"Visual Question Answering"},{"task_slug":"visual-question-answering","task_name":"Visual Question Answering (VQA)"}],"methods":[{"method_slug":"focus","method_name":"Focus"}],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/visual-question-answering-on-v-bench","task":"Visual Question Answering","dataset":"V*bench","model":"LLaVA-OneVision7B w. FOCUS","rank_in_archive_order":1,"of":5,"metrics":{"Accuracy":"92.15"},"uses_additional_data":false}],"syntology":{"atlas_url":null,"mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}