{"id":"https://openalex.org/W7203483625","doi":"https://doi.org/10.48550/arxiv.2608.13141","title":"MergeOver: Post-Training Token Merging for Recursive Vision Transformers","display_name":"MergeOver: Post-Training Token Merging for Recursive Vision Transformers","publication_year":2026,"publication_date":"2026-08-13","ids":{"openalex":"https://openalex.org/W7203483625","doi":"https://doi.org/10.48550/arxiv.2608.13141"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2608.13141","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.13141","pdf_url":"https://arxiv.org/pdf/2608.13141","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2608.13141","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5147157991","display_name":"Junseo Kim","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I94624287","display_name":"University of Twente","ror":"https://ror.org/006hf6230","country_code":"NL","type":"education","lineage":["https://openalex.org/I94624287"]}],"countries":["NL"],"is_corresponding":false,"raw_author_name":"Kim, Junseo","raw_affiliation_strings":["Computer Architecture for Embedded Systems , University of Twente Enschede , The Netherlands"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Computer Architecture for Embedded Systems , University of Twente Enschede , The Netherlands","institution_ids":["https://openalex.org/I94624287"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5024917187","display_name":"Uraz Odyurt","orcid":"https://orcid.org/0000-0003-1094-0234","observed_orcids":["https://orcid.org/0000-0003-1094-0234"]},"institutions":[{"id":"https://openalex.org/I94624287","display_name":"University of Twente","ror":"https://ror.org/006hf6230","country_code":"NL","type":"education","lineage":["https://openalex.org/I94624287"]}],"countries":["NL"],"is_corresponding":false,"raw_author_name":"Odyurt, Uraz","raw_affiliation_strings":["Faculty of Engineering Technology , University of Twente Enschede , The Netherlands"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Faculty of Engineering Technology , University of Twente Enschede , The Netherlands","institution_ids":["https://openalex.org/I94624287"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5040160722","display_name":"Amirreza Yousefzadeh","orcid":"https://orcid.org/0000-0002-2967-5090","observed_orcids":["https://orcid.org/0000-0002-2967-5090"]},"institutions":[{"id":"https://openalex.org/I94624287","display_name":"University of Twente","ror":"https://ror.org/006hf6230","country_code":"NL","type":"education","lineage":["https://openalex.org/I94624287"]}],"countries":["NL"],"is_corresponding":false,"raw_author_name":"Yousefzadeh, Amirreza","raw_affiliation_strings":["Computer Architecture for Embedded Systems , University of Twente Enschede , The Netherlands"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Computer Architecture for Embedded Systems , University of Twente Enschede , The Netherlands","institution_ids":["https://openalex.org/I94624287"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I94624287"],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.9994999766349792,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.9994999766349792,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.00019999999494757503,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T11605","display_name":"Visual Attention and Saliency Detection","score":9.999999747378752e-05,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/token-merging","display_name":"token merging","score":0.8889999985694885},{"id":"https://openalex.org/keywords/post-training-optimization","display_name":"post-training optimization","score":0.5400000214576721},{"id":"https://openalex.org/keywords/parameter-reduction","display_name":"parameter reduction","score":0.49300000071525574},{"id":"https://openalex.org/keywords/edge-deployment","display_name":"edge deployment","score":0.8669999837875366},{"id":"https://openalex.org/keywords/hierarchical-vision-transformer","display_name":"hierarchical vision transformer","score":0.4580000042915344}],"concepts":[{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.7324000000953674},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7153000235557556},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.45410001277923584},{"id":"https://openalex.org/C79403827","wikidata":"https://www.wikidata.org/wiki/Q3988","display_name":"Real-time computing","level":1,"score":0.4083000123500824},{"id":"https://openalex.org/C82876162","wikidata":"https://www.wikidata.org/wiki/Q17096504","display_name":"Latency (audio)","level":2,"score":0.382999986410141},{"id":"https://openalex.org/C118505674","wikidata":"https://www.wikidata.org/wiki/Q42586063","display_name":"Encoder","level":2,"score":0.35010001063346863},{"id":"https://openalex.org/C206729178","wikidata":"https://www.wikidata.org/wiki/Q2271896","display_name":"Scheduling (production processes)","level":2,"score":0.3418000042438507},{"id":"https://openalex.org/C179799912","wikidata":"https://www.wikidata.org/wiki/Q205084","display_name":"Computational complexity theory","level":2,"score":0.336899995803833},{"id":"https://openalex.org/C129844170","wikidata":"https://www.wikidata.org/wiki/Q41299","display_name":"Quadratic equation","level":2,"score":0.33090001344680786},{"id":"https://openalex.org/C68387754","wikidata":"https://www.wikidata.org/wiki/Q7271585","display_name":"Schedule","level":2,"score":0.3172000050544739},{"id":"https://openalex.org/C175291020","wikidata":"https://www.wikidata.org/wiki/Q1156822","display_name":"Offset (computer science)","level":2,"score":0.31220000982284546},{"id":"https://openalex.org/C157764524","wikidata":"https://www.wikidata.org/wiki/Q1383412","display_name":"Throughput","level":3,"score":0.2937000095844269},{"id":"https://openalex.org/C111335779","wikidata":"https://www.wikidata.org/wiki/Q3454686","display_name":"Reduction (mathematics)","level":2,"score":0.27250000834465027},{"id":"https://openalex.org/C147764199","wikidata":"https://www.wikidata.org/wiki/Q6865248","display_name":"Minification","level":2,"score":0.27149999141693115},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.2542000114917755},{"id":"https://openalex.org/C138236772","wikidata":"https://www.wikidata.org/wiki/Q25098575","display_name":"Edge device","level":3,"score":0.2531999945640564},{"id":"https://openalex.org/C113954288","wikidata":"https://www.wikidata.org/wiki/Q186885","display_name":"Timestamp","level":2,"score":0.2513999938964844}],"mesh":[],"locations_count":3,"locations":[{"id":"pmh:oai:arXiv.org:2608.13141","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.13141","pdf_url":"https://arxiv.org/pdf/2608.13141","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"pmh:oai:ris.utwente.nl:openaire/6ba6cd09-8b55-42cd-a4a4-9b9dfcead263","doi":null,"is_oa":true,"landing_page_url":"https://research.utwente.nl/en/publications/6ba6cd09-8b55-42cd-a4a4-9b9dfcead263","pdf_url":null,"source":{"id":"https://openalex.org/S4406922991","display_name":"University of Twente Research Information","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":"Kim, J, Odyurt, U & Yousefzadeh, A 2026 'MergeOver : Post-Training Token Merging for Recursive Vision Transformers' ArXiv.org. https://doi.org/10.48550/arXiv.2608.13141","raw_type":"info:eu-repo/semantics/preprint"},{"id":"doi:10.48550/arxiv.2608.13141","doi":"https://doi.org/10.48550/arxiv.2608.13141","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2608.13141","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2608.13141","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.13141","pdf_url":"https://arxiv.org/pdf/2608.13141","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7203483625.pdf","grobid_xml":"https://content.openalex.org/works/W7203483625.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7154480968","https://openalex.org/W7203830332","https://openalex.org/W4394593136","https://openalex.org/W4323556158","https://openalex.org/W7162635299","https://openalex.org/W4306886919","https://openalex.org/W4406631984","https://openalex.org/W4385004455","https://openalex.org/W4415428809","https://openalex.org/W4384159680"],"abstract_inverted_index":{"Vision":[0],"Transformers":[1],"(ViTs)":[2],"demonstrate":[3],"exceptional":[4],"performance":[5],"in":[6],"computer":[7],"vision":[8],"but":[9,168],"suffer":[10],"from":[11],"large":[12],"parameter":[13,31],"counts":[14,32],"and":[15,33,38,82,93,117,151,157,188,194,209,218],"quadratic":[16],"computational":[17,37],"complexity,":[18],"severely":[19],"limiting":[20],"their":[21],"deployment":[22],"on":[23,129],"resource-constrained":[24],"edge":[25],"hardware.":[26],"While":[27],"recursive":[28,126,213,228],"weight-sharing":[29,214],"reduces":[30,134,145,184],"token":[34,108,224],"merging":[35,94,225],"mitigates":[36],"memory":[39,148,210],"bottlenecks,":[40],"integrating":[41],"these":[42],"two":[43],"paradigms":[44],"without":[45,216],"costly":[46],"retraining":[47],"is":[48],"non-trivial,":[49],"leaving":[50],"this":[51,97],"intersection":[52],"largely":[53],"unexplored.":[54],"We":[55,99],"propose":[56],"MergeOver,":[57],"a":[58,102,119,177,203,220],"post-training":[59],"approach":[60],"that":[61,106,199,212],"integrates":[62],"Token":[63],"Merging":[64],"(ToMe)":[65],"into":[66],"the":[67,91,111,142,207],"recursively":[68],"weight-shared":[69],"Sliced":[70],"Recursive":[71],"Transformer":[72],"(SReT).":[73],"Through":[74],"an":[75],"Unmerge":[76],"tracking":[77,85],"stack,":[78],"constraint-safe":[79],"merge-rate":[80],"adjustment,":[81],"synchronised":[83],"token-mass":[84],"across":[86],"spatial":[87,92],"permutations,":[88],"MergeOver":[89,200],"resolves":[90],"constraints":[95],"of":[96,114,206],"integration.":[98],"further":[100],"employ":[101],"stage-wise":[103],"single-shot":[104],"schedule":[105],"performs":[107],"reduction":[109],"at":[110,153,164,172,190],"first":[112],"block":[113],"each":[115],"stage":[116],"maintains":[118],"fixed":[120],"sequence":[121],"length":[122],"throughout":[123],"its":[124],"subsequent":[125],"iterations.":[127],"Benchmarked":[128],"ImageNet-1K,":[130],"our":[131],"selected":[132],"configuration":[133],"top-1":[135],"accuracy":[136],"by":[137,149,162,170,186],"1.47":[138],"percentage":[139],"points.":[140],"On":[141,176],"GPU,":[143],"it":[144,183],"peak":[146],"activation":[147],"37.3%":[150],"38.4%":[152],"batch":[154,165,173,191],"sizes":[155,192],"1":[156,167,193],"16,":[158],"while":[159],"throughput":[160,208],"decreases":[161],"21.7%":[163,171],"size":[166,174],"increases":[169],"16.":[175,195],"Raspberry":[178],"Pi":[179],"5":[180],"(ARM":[181],"CPU),":[182],"latency":[185],"2.4%":[187],"17.6%":[189],"These":[196],"results":[197],"show":[198],"can":[201],"recover":[202],"meaningful":[204],"part":[205],"cost":[211],"introduces,":[215],"retraining,":[217],"provides":[219],"baseline":[221],"for":[222],"combining":[223],"with":[226],"hierarchical":[227],"transformers.":[229]},"counts_by_year":[],"updated_date":"2026-09-30T09:27:27.532063","created_date":"2026-08-15T00:00:00"}