{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2025,10,15]],"date-time":"2025-10-15T14:12:54Z","timestamp":1760537574718,"version":"build-2065373602"},"reference-count":0,"publisher":"Association for the Advancement of Artificial Intelligence (AAAI)","issue":"3","content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":["AIES"],"abstract":"<jats:p>ColPali recently proposed a method for explaining multimodal retrieval-augmented generation (RAG) by visualizing how vision\u2013language models (VLMs) connect image patches to text tokens. \nHowever, our theoretical analysis and experiments show that these similarity-based saliency maps are fragile and often misleading.\nWe therefore caution against relying solely on intuitive visualizations and present a principled patch-level dissection technique that traces how vision LLMs actually accumulate evidence across modalities.\nTo address this issue, we introduce Needle-in-a-Patched-Haystack: a patch-centered dataset and metric suite that quantifies transparency by benchmarking localization performance in vision LLMs. Together, our analysis and toolkit establish a stricter standard for VLM interpretability and provide a drop-in evaluation protocol for future research on robust, multimodal explanations.<\/jats:p>","DOI":"10.1609\/aies.v8i3.36763","type":"journal-article","created":{"date-parts":[[2025,10,15]],"date-time":"2025-10-15T13:31:33Z","timestamp":1760535093000},"page":"2832-2839","source":"Crossref","is-referenced-by-count":0,"title":["Needle in a Patched Haystack: Evaluating Saliency Maps for Vision LLMs"],"prefix":"10.1609","volume":"8","author":[{"given":"Bastien","family":"Zimmermann","sequence":"first","affiliation":[],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Matthieu","family":"Boussard","sequence":"additional","affiliation":[],"role":[{"role":"author","vocabulary":"crossref"}]}],"member":"9382","published-online":{"date-parts":[[2025,10,15]]},"container-title":["Proceedings of the AAAI\/ACM Conference on AI, Ethics, and Society"],"original-title":[],"link":[{"URL":"https:\/\/ojs.aaai.org\/index.php\/AIES\/article\/download\/36763\/38901","content-type":"application\/pdf","content-version":"vor","intended-application":"text-mining"},{"URL":"https:\/\/ojs.aaai.org\/index.php\/AIES\/article\/download\/36763\/38901","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2025,10,15]],"date-time":"2025-10-15T13:31:33Z","timestamp":1760535093000},"score":1,"resource":{"primary":{"URL":"https:\/\/ojs.aaai.org\/index.php\/AIES\/article\/view\/36763"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2025,10,15]]},"references-count":0,"journal-issue":{"issue":"3","published-online":{"date-parts":[[2025,10,15]]}},"URL":"https:\/\/doi.org\/10.1609\/aies.v8i3.36763","relation":{},"ISSN":["3065-8365"],"issn-type":[{"value":"3065-8365","type":"electronic"}],"subject":[],"published":{"date-parts":[[2025,10,15]]}}}