[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-81851-en":3,"doc-seo-81851-105":31,"detail-sidebar-cat-0-en-105":93},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":21,"is_downloadable":21,"audit_status":21,"page_count":22,"language":23,"language_code":24,"site_id":25,"html_lang":24,"table_of_contents":26,"faqs":27,"seo_title":28,"seo_description":14,"update_tm":29,"read_time":30},81851,5909877438554,"Maeve","https://ap-avatar.wpscdn.com/avatar/5600025385ad2bf12a7?_k=1778553567797529272",8,"Research & Report","Fusion A Framework for Unified Sequential Token Adaptation in Vision Transformers","Vision Transformers deliver strong image classification accuracy but waste compute by processing all patch tokens through every transformer block, even when many tokens are redundant or uninformative. Prior adaptive inference methods—token pruning, token merging, and early exiting—can destabilize intermediate representations when combined naively. Fusion proposes a unified sequential adaptive inference pipeline, merging before confidence evaluation and pruning only for samples that continue. Lightweight routers and profile-aware scaling enable input-adaptive compression and inference-time accuracy–latency adjustment without retraining. On ImageNet-1k with DeiT-S, Fusion matches or exceeds state-of-the-art while reducing calibration error up to 4× and inference energy by 48%, with consistent cross-dataset transfer.","© 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. This paper is accepted at the 29th Euromicro Conference Series on Digital System Design (DSD) 2026 .   \nFusion: A Framework for Unified Sequential Token AdaptatIon in VisiOn TraNsformers  \nAravind Pradeep 1 , Samira Nazari2 , Mahdi Taheri3,4 , and Christian Herglotz 1  \n1Brandenburg University of Technology Cottbus-Senftenberg, Germany  \n2University of Zanjan, Iran  \n3Humboldt University, Berlin, Germany  \n4Tallinn University of Technology, Tallinn, Estonia  \narXiv :2607 .026 12v 1 [ cs .CV] 1 Jul 2026  \nAbstract—Vision Transformers achieve strong image classification accuracy but process all image regions with nearly the same computation, even when many regions are redundant or uninformative. Recent adaptive inference methods reduce this cost by selectively compressing tokens or terminating inference early, but combining these mechanisms often causes unstable intermediate representations and accuracy degradation. We introduce Fusion, a unified adaptive inference framework that coordinates token merging, early exiting, and token pruning through a simplestaged design: tokens are merged first, confidence is evaluated next, and pruning is applied only to samples that continue inference. This ordering allows the three mechanisms to operate cooperatively rather than competitively. Fusion further includes lightweight routing modules that adapt compression strength to each input and support inference-time adjustment of the accuracy–latency trade-off without retraining. On ImageNet- 1k with DeiT-S, Fusion matches or surpasses state-of-the-art adaptive ViT methods at comparable compute budgets while reducing calibration error by up to 4× and inference energy by 48%. Experiments across ImageNet-100, CIFAR-100, and ImageNette with multiple ViT backbones demonstrate consistent transferability without dataset-specific tuning.  \nIndex Terms—Vision Transformer, Token Pruning, Early Exit, Token Merging, Adaptive Inference, Efficient Deep Learning  \nI. INTRODUCTION  \nVision Transformers (ViTs) [1], [2] achieve strong image classification accuracy but incur substantial computational cost due to uniform processing of all tokens across all transformer blocks. A 224 × 224 image produces 196 patch tokens that traverse all twelve layers of a DeiT-S backbone [2] (a 22Mparameter data-efficient Vision Transformer), regardless of whether the input image is simple or highly structured. This design leads to significant redundancy during inference.  \nDynamic token processing methods address this inefficiency along three complementary axes: token pruning [3], [4] removes uninformative tokens, token merging [5] aggregates redundant tokens, and early exit [6], [7], [8] reduces the number of transformer blocks evaluated for a given input, terminating inference once the prediction is sufficiently confident. These mechanisms are typically developed and evaluated independently, each treating the ViT backbone as a fixed computation graph.  \nHowever, combining these mechanisms does not yield additive efficiency gains. When pruning, merging, and early exit  \nare applied simultaneously as independently trained modules, their adaptive decisions interfere through shared intermediate representations (the per-block token embeddings produced inside the transformer) . On ImageNet-1k with DeiT-S, naïve parallel composition reduces accuracy by 1.62%, relative to the baseline and by 1.58%, relative to a sequential composition of the same mechanisms (Table VII) . The earlyexit head suffers the most: at the 8th transformer block, its standalone accuracy drops from 89.4% to 47.6% under parallel couplin","cbCaiolcElUiCKtz","https://ap.wps.com/l/cbCaiolcElUiCKtz","pdf",1966320,5,1,9,"English","en",105,"# I. INTRODUCTION\n# II. RELATED WORK\n# III. PROPOSED FRAMEWORK\n# IV. EXPERIMENTAL RESULTS AND ABLATION STUDIES","[{\"question\":\"Why do Vision Transformers waste computation during inference?\",\"answer\":\"ViTs process all patch tokens through each transformer block, even when many tokens are redundant or uninformative for the final prediction.\"},{\"question\":\"What problem occurs when token pruning, merging, and early exit are combined in a naive parallel way?\",\"answer\":\"Their independently trained decisions interfere through shared intermediate representations, causing accuracy degradation—most severely impacting the early-exit head.\"},{\"question\":\"How does Fusion coordinate the three adaptive mechanisms to improve stability?\",\"answer\":\"Fusion uses a staged sequential pipeline: merge first, then evaluate confidence for early exiting, and apply pruning only to samples that continue inference.\"}]","Fusion A Framework for Unified Sequential Token Adaptation in Vision Transformers | PDF",1784176641,23,{"code":4,"msg":32,"data":33},"ok",{"site_id":25,"language":24,"slug":34,"title":13,"keywords":35,"description":14,"schema_data":36,"social_meta":88,"head_meta":90,"extra_data":92,"updated_unix":29},"fusion-a-framework-for-unified-sequential-token-adaptation-in-vision-transformers","",{"@graph":37,"@context":87},[38,55,70],{"@type":39,"itemListElement":40},"BreadcrumbList",[41,45,49,52],{"item":42,"name":43,"@type":44,"position":21},"https://docshare.wps.com","Home","ListItem",{"item":46,"name":47,"@type":44,"position":48},"https://docshare.wps.com/document/","Document",2,{"item":50,"name":12,"@type":44,"position":51},"https://docshare.wps.com/document/research-report/",3,{"item":53,"name":13,"@type":44,"position":54},"https://docshare.wps.com/document/fusion-a-framework-for-unified-sequential-token-adaptation-in-vision-transformers/81851/",4,{"url":53,"name":13,"@type":56,"author":57,"headline":13,"publisher":59,"fileFormat":62,"inLanguage":24,"description":14,"dateModified":63,"datePublished":64,"encodingFormat":62,"isAccessibleForFree":65,"interactionStatistic":66},"DigitalDocument",{"name":9,"@type":58},"Person",{"url":42,"name":60,"@type":61},"DocShare","Organization","application/pdf","2026-07-29","2026-07-16",true,{"@type":67,"interactionType":68,"userInteractionCount":20},"InteractionCounter",{"@type":69},"ViewAction",{"@type":71,"mainEntity":72},"FAQPage",[73,79,83],{"name":74,"@type":75,"acceptedAnswer":76},"Why do Vision Transformers waste computation during inference?","Question",{"text":77,"@type":78},"ViTs process all patch tokens through each transformer block, even when many tokens are redundant or uninformative for the final prediction.","Answer",{"name":80,"@type":75,"acceptedAnswer":81},"What problem occurs when token pruning, merging, and early exit are combined in a naive parallel way?",{"text":82,"@type":78},"Their independently trained decisions interfere through shared intermediate representations, causing accuracy degradation—most severely impacting the early-exit head.",{"name":84,"@type":75,"acceptedAnswer":85},"How does Fusion coordinate the three adaptive mechanisms to improve stability?",{"text":86,"@type":78},"Fusion uses a staged sequential pipeline: merge first, then evaluate confidence for early exiting, and apply pruning only to samples that continue inference.","https://schema.org",{"og:url":53,"og:type":89,"og:title":13,"og:site_name":60,"og:description":14},"article",{"robots":91,"canonical":53},"index,follow",{"doc_id":7,"site_id":25},{"code":4,"msg":5,"data":94},[95,99,103,107,111,116,121,124,128,131,135],{"id":21,"doc_module":4,"doc_module_name":47,"category_name":96,"show_sort_weight":97,"slug":98},"Story & Novel",90,"story-novel",{"id":48,"doc_module":4,"doc_module_name":47,"category_name":100,"show_sort_weight":101,"slug":102},"Literature",80,"literature",{"id":54,"doc_module":4,"doc_module_name":47,"category_name":104,"show_sort_weight":105,"slug":106},"Exam",70,"exam",{"id":20,"doc_module":4,"doc_module_name":47,"category_name":108,"show_sort_weight":109,"slug":110},"Comic",60,"comic",{"id":112,"doc_module":4,"doc_module_name":47,"category_name":113,"show_sort_weight":114,"slug":115},6,"Technology",50,"technology",{"id":117,"doc_module":4,"doc_module_name":47,"category_name":118,"show_sort_weight":119,"slug":120},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":47,"category_name":12,"show_sort_weight":122,"slug":123},30,"research-report",{"id":22,"doc_module":4,"doc_module_name":47,"category_name":125,"show_sort_weight":126,"slug":127},"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":47,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":47,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":47,"category_name":137,"show_sort_weight":20,"slug":138},19,"General","general"]