{"id":"https://openalex.org/W7162670443","doi":"https://doi.org/10.48550/arxiv.2605.28051","title":"Beyond Surrogate Gradients: Fully Differentiable Token Pruning for Vision-Language Models","display_name":"Beyond Surrogate Gradients: Fully Differentiable Token Pruning for Vision-Language Models","publication_year":2026,"publication_date":"2026-05-27","ids":{"openalex":"https://openalex.org/W7162670443","doi":"https://doi.org/10.48550/arxiv.2605.28051"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.28051","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.28051","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2605.28051","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5128646065","display_name":"Landi He","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"He, Landi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5002617843","display_name":"Mingde Yao","orcid":"https://orcid.org/0000-0001-5994-6288"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yao, Mingde","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5127926984","display_name":"Shawn Young","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Young, Shawn","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5137203839","display_name":"Lijian Xu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xu, Lijian","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":false,"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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.9763000011444092,"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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.9763000011444092,"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":0.005400000140070915,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.0020000000949949026,"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/security-token","display_name":"Security token","score":0.7516000270843506},{"id":"https://openalex.org/keywords/pruning","display_name":"Pruning","score":0.7064999938011169},{"id":"https://openalex.org/keywords/selection","display_name":"Selection (genetic algorithm)","score":0.5662999749183655},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.5235000252723694},{"id":"https://openalex.org/keywords/differentiable-function","display_name":"Differentiable function","score":0.5198000073432922},{"id":"https://openalex.org/keywords/path","display_name":"Path (computing)","score":0.45910000801086426}],"concepts":[{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.7516000270843506},{"id":"https://openalex.org/C108010975","wikidata":"https://www.wikidata.org/wiki/Q500094","display_name":"Pruning","level":2,"score":0.7064999938011169},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6942999958992004},{"id":"https://openalex.org/C81917197","wikidata":"https://www.wikidata.org/wiki/Q628760","display_name":"Selection (genetic algorithm)","level":2,"score":0.5662999749183655},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5235000252723694},{"id":"https://openalex.org/C202615002","wikidata":"https://www.wikidata.org/wiki/Q783507","display_name":"Differentiable function","level":2,"score":0.5198000073432922},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4950000047683716},{"id":"https://openalex.org/C2777735758","wikidata":"https://www.wikidata.org/wiki/Q817765","display_name":"Path (computing)","level":2,"score":0.45910000801086426},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.3285999894142151},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.32749998569488525},{"id":"https://openalex.org/C115067241","wikidata":"https://www.wikidata.org/wiki/Q1639854","display_name":"Token passing","level":3,"score":0.32510000467300415},{"id":"https://openalex.org/C2775924081","wikidata":"https://www.wikidata.org/wiki/Q55608371","display_name":"Control (management)","level":2,"score":0.3188999891281128},{"id":"https://openalex.org/C137836250","wikidata":"https://www.wikidata.org/wiki/Q984063","display_name":"Optimization problem","level":2,"score":0.2847000062465668},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.2824999988079071},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.27399998903274536},{"id":"https://openalex.org/C185429906","wikidata":"https://www.wikidata.org/wiki/Q1130160","display_name":"Estimator","level":2,"score":0.2703000009059906},{"id":"https://openalex.org/C179799912","wikidata":"https://www.wikidata.org/wiki/Q205084","display_name":"Computational complexity theory","level":2,"score":0.26739999651908875},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.26260000467300415}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.28051","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.28051","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2605.28051","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.28051","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Visual":[0],"token":[1,47,62,78,100,111],"pruning":[2,57],"reduces":[3],"the":[4,29,38,122],"computational":[5],"cost":[6],"of":[7,46,61,65,132,145],"Vision-Language":[8],"Models":[9],"(VLMs)":[10],"by":[11,33,139],"removing":[12],"redundant":[13],"visual":[14],"tokens.":[15],"Existing":[16],"methods":[17],"typically":[18],"rely":[19],"on":[20,83,99,121],"Gumbel-Softmax":[21],"to":[22,43],"approximate":[23],"discrete":[24,66],"selection":[25,40,67],"during":[26,93],"training.":[27,94],"However,":[28],"optimization":[30,107],"is":[31],"driven":[32],"surrogate":[34],"gradients":[35],"rather":[36],"than":[37],"true":[39],"process,":[41],"leading":[42],"unreliable":[44],"learning":[45,110],"importance.":[48,112],"In":[49],"this":[50],"paper,":[51],"we":[52,70],"propose":[53],"DiffPrune,":[54],"which":[55],"reformulates":[56],"as":[58],"continuous":[59],"control":[60],"information":[63,91],"instead":[64],"learning.":[68],"Specifically,":[69],"introduce":[71],"an":[72],"Information":[73],"Throttler":[74],"that":[75],"modulates":[76],"each":[77],"using":[79],"variance-preserving":[80],"noise":[81],"conditioned":[82],"importance":[84],"scores,":[85],"where":[86],"higher":[87],"scores":[88],"induce":[89],"less":[90],"suppression":[92],"This":[95],"design":[96],"directly":[97],"operates":[98],"representations,":[101],"naturally":[102],"providing":[103],"a":[104],"fully":[105],"differentiable":[106],"path":[108],"for":[109],"At":[113],"inference,":[114],"tokens":[115],"are":[116],"removed":[117],"via":[118],"hard":[119],"thresholding":[120],"learned":[123],"scores.":[124],"Across":[125],"ten":[126],"VLM":[127],"benchmarks,":[128],"DiffPrune":[129],"retains":[130],"96.5%":[131],"full-model":[133],"accuracy":[134],"while":[135],"accelerating":[136],"LLM":[137],"prefill":[138],"2.85x,":[140],"with":[141],"only":[142],"0.69":[143],"ms":[144],"inference":[146],"overhead.":[147]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-05-29T00:00:00"}