[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-328131-105":59,"doc-detail-328131-en":131},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":124,"head_meta":126,"extra_data":128,"updated_unix":130},105,"en","propose-and-attend-multi-token-localized-attention-mtla","Propose and Attend - Multi-Token Localized Attention (MTLA)","","Multimodal large language models output localized predictions like bounding boxes and temporal windows, yet frequently hallucinate these regions. The model’s token log-probabilities provide weak guidance because they blur grounding quality with input ambiguity and become near-deterministic after commitment. MultiToken Localized Attention (MTLA) introduces a training-free posthoc score that measures how strongly each prediction token attends to the region it claims, improving grounding reliability. Across multiple MLLM families and three modalities, MTLA boosts hallucination AUROC by +7 to +21 over prior training-free baselines and nearly doubles zero-shot COCO detection AP for an open-source 8B model without task-specific training.",{"@graph":69,"@context":123},[70,84,106],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/propose-and-attend-multi-token-localized-attention-mtla/328131/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/propose-and-attend-multi-token-localized-attention-mtla/328131.png","ImageObject",300,407,{"name":92,"@type":93},"Evangeline","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-23","2026-09-21",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",13,{"@type":107,"mainEntity":108},"FAQPage",[109,115,119],{"name":110,"@type":111,"acceptedAnswer":112},"Why do multimodal large language models often produce hallucinated localization regions?","Question",{"text":113,"@type":114},"They emit localized predictions through language generation, but a large fraction of emitted regions do not match valid objects or events in the input, indicating a reliability gap.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"What is MultiToken Localized Attention (MTLA) and how does it compute confidence?",{"text":118,"@type":114},"MTLA is a training-free, posthoc score that measures how strongly a prediction’s tokens attend to the region the model claims, using restricted attention mass within the proposed region.",{"name":120,"@type":111,"acceptedAnswer":121},"How effective is MTLA compared with prior training-free baselines?",{"text":122,"@type":114},"Across multiple MLLM families and three modalities, MTLA improves hallucination AUROC by +7 to +21 over the best prior training-free baseline and boosts localization accuracy in image, video, and audio.","https://schema.org",{"og:url":83,"og:type":125,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":127,"canonical":83},"index,follow",{"doc_id":129,"site_id":62},328131,1790195073,{"code":4,"msg":5,"data":132},{"doc_id":129,"user_id":133,"nickname":92,"user_avatar":134,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":135,"file_id":136,"file_url":137,"file_type":138,"file_size":139,"view_count":105,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":140,"language":141,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":142,"faqs":143,"seo_title":144,"seo_description":67,"update_tm":145,"read_time":146,"preview_page_count":4},13056703019662,"https://ap-avatar.wpscdn.com/avatar/be000253a8e92610077?_k=1778726343310543188","Propose and Attend: Training-free MLLM Grounding Confidence via  \nMulti-Token Localized Attention  \nDaniel Shalam Emanuel Ben Baruch Avi Ben Cohen Tal Remez  \nAmazon  \nCode: [https://github.com/TalRemez/MTLA.git](https://github.com/TalRemez/MTLA.git)  \narXiv :2607 .05978v 1 [ cs .CV] 7 Jul 2026  \nAbstract  \nMultimodal large language models can emit localized predictions, bounding boxes for objects and temporal windows for video and audio events, but they hallucinate these regions prolifically. The model’s own token log-probabilities are nearly uninformative: they conflate grounding quality with input ambiguity, and coordinate tokens become neardeterministic once the model commits. We propose MultiToken Localized Attention (MTLA): a training-free, posthoc score that measures how strongly a prediction’s tokens attend to the region they claim. Prior attention-based detectors, which sum attention over the entire input modality and read a single response token, are weaker special cases; we show that summing only within the claimed region and aggregating across all prediction tokens recoversa stronger grounding signal. The same recipe applies almost trivially to other modalities and tasks: object detection in images and temporal localization in video and audio. Across multiple MLLM families and three modalities, MTLA improves hallucination AUROC by +7 to +21 over the best prior training-free baseline. Used as a confidence score for re-ranking, it nearly doubles the zero-shot COCO detection AP of an open-source 8B generalist (from 20.4 to 37.0), narrowing the gap to supervised detectors without any task-specific training.  \n1. Introduction  \nRecent multimodal large language models (MLLMs), including Qwen [1], Gemma 4 [11], and Audio Flamingo 3 [13], are capable of producing localization outputs, such as spatial bounding boxes and temporal windows, directly through language generation. Trained on largescale grounding corpora such as RefCOCO [37] and Visual Genome [20], these models can localize objects, events, and activities without dedicated task-specific heads. The same generation mechanism naturally extends across modalities, enabling object localization in images, temporal localiza-  \nAP  \n40  \n30  \n20  \nImage (COCO)  42.0  \n37.0  \nQwen3-VL-8B  \nMTLA (ours)  \nDETR  \nAP  \n40  \n30  \nVideo (QVHighlights)  39.9  \n36.6  \nPSDS1  \n0.35  \n0.30  \n0.25  \n0.20  \nQwen3-VL-8B  \nMTLA (ours)  \nQD-DETR Audio Flamingo 3  \nMTLA (ours)  \nBEATs  \n Zero-shot baseline  MTLA (ours)  Supervised  \nFigure 1 . MTLA enables reliable localization from MLLMs. Top: MLLM detection outputs contain many hallucinated predictions (red); MTLA uses the model’s own attention to estimate localization confidence and suppress them, retaining grounded predictions (blue) . Bottom: applied post hoc, MTLA (N=16 selfconsistency) lifts each base MLLM’s localization accuracy across image, video, and audio, narrowing the gap to supervised specialists.  \ntion in video, and sound-event localization in audio.  \nLocalization has traditionally been addressed using specialized architectures. Closed-vocabulary detectors such as DETR [2] and open-vocabulary systems such as GLIP [23], OWL-ViT [29], and Grounding DINO [26] are explicitly designed and optimized for localization tasks. In contrast, MLLMs provide a unified framework for localization, allowing a single model to operate across modalities and query types without task-specific architectures or ad hoc training. As these models continue to improve, an important question emerges: can MLLMs serve as reliable localization systems?  \nDespite their growing use for localization tasks, MLLMs are rarely evaluated using the standard metrics of the de-  \nFigure 2 . Multi-Token Localized Attention (MTLA). A pre-trained MLLM localizes objects, emitting per prediction a token sequence of box coordinates and a class label. We read the decoder’s attention from these prediction tokens Qp onto the input, then restrict it to the patches inside the ","cbCaicEm30Zae8N3","https://ap.wps.com/l/cbCaicEm30Zae8N3","pdf",30226424,15,"English","# Abstract\n# 1. Introduction\n# Localization reliability gap\n# Confidence estimation limitations\n# Related work on hallucination detection","[{\"question\":\"Why do multimodal large language models often produce hallucinated localization regions?\",\"answer\":\"They emit localized predictions through language generation, but a large fraction of emitted regions do not match valid objects or events in the input, indicating a reliability gap.\"},{\"question\":\"What is MultiToken Localized Attention (MTLA) and how does it compute confidence?\",\"answer\":\"MTLA is a training-free, posthoc score that measures how strongly a prediction’s tokens attend to the region the model claims, using restricted attention mass within the proposed region.\"},{\"question\":\"How effective is MTLA compared with prior training-free baselines?\",\"answer\":\"Across multiple MLLM families and three modalities, MTLA improves hallucination AUROC by +7 to +21 over the best prior training-free baseline and boosts localization accuracy in image, video, and audio.\"}]","Propose and Attend - Multi-Token Localized Attention (MTLA) | PDF",1789978714,38]