{"id":"https://openalex.org/W7212059249","doi":"https://doi.org/10.48550/arxiv.2609.07264","title":"Dense Structural Compression of Transformers via Gauge-Correct Channel Removal","display_name":"Dense Structural Compression of Transformers via Gauge-Correct Channel Removal","publication_year":2026,"publication_date":"2026-09-07","ids":{"openalex":"https://openalex.org/W7212059249","doi":"https://doi.org/10.48550/arxiv.2609.07264"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2609.07264","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.07264","pdf_url":"https://arxiv.org/pdf/2609.07264","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":null,"license_id":null,"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/2609.07264","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5031751484","display_name":"Jed A. Duersch","orcid":"https://orcid.org/0000-0002-2178-3695","observed_orcids":["https://orcid.org/0000-0002-2178-3695"]},"institutions":[{"id":"https://openalex.org/I1294671590","display_name":"Centre National de la Recherche Scientifique","ror":"https://ror.org/02feahw73","country_code":"FR","type":"government","lineage":["https://openalex.org/I1294671590"]},{"id":"https://openalex.org/I4405277525","display_name":"Centre de Recherche en Informatique de Lens","ror":"https://ror.org/03072md54","country_code":"FR","type":"facility","lineage":["https://openalex.org/I1294671590","https://openalex.org/I4405277525","https://openalex.org/I44563897"]},{"id":"https://openalex.org/I44563897","display_name":"Universit\u00e9 d'Artois","ror":"https://ror.org/053x9s498","country_code":"FR","type":"education","lineage":["https://openalex.org/I44563897"]}],"countries":["FR"],"is_corresponding":false,"raw_author_name":"Duersch, Jed A.","raw_affiliation_strings":["Universit\u00e9 d'Artois , CNRS , CRIL UMR 8188 , Lens , France"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Universit\u00e9 d'Artois , CNRS , CRIL UMR 8188 , Lens , France","institution_ids":["https://openalex.org/I1294671590","https://openalex.org/I4405277525","https://openalex.org/I44563897"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5107179437","display_name":"Na\u00efm Es-Sebbani","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I1294671590","display_name":"Centre National de la Recherche Scientifique","ror":"https://ror.org/02feahw73","country_code":"FR","type":"government","lineage":["https://openalex.org/I1294671590"]},{"id":"https://openalex.org/I4405277525","display_name":"Centre de Recherche en Informatique de Lens","ror":"https://ror.org/03072md54","country_code":"FR","type":"facility","lineage":["https://openalex.org/I1294671590","https://openalex.org/I4405277525","https://openalex.org/I44563897"]},{"id":"https://openalex.org/I44563897","display_name":"Universit\u00e9 d'Artois","ror":"https://ror.org/053x9s498","country_code":"FR","type":"education","lineage":["https://openalex.org/I44563897"]}],"countries":["FR"],"is_corresponding":false,"raw_author_name":"Es-Sebbani, Na\u00efm","raw_affiliation_strings":["Universit\u00e9 d'Artois , CNRS , CRIL UMR 8188 , Lens , France"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Universit\u00e9 d'Artois , CNRS , CRIL UMR 8188 , Lens , France","institution_ids":["https://openalex.org/I1294671590","https://openalex.org/I4405277525","https://openalex.org/I44563897"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5126101223","display_name":"Nathana\u00ebl Haas","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I1294671590","display_name":"Centre National de la Recherche Scientifique","ror":"https://ror.org/02feahw73","country_code":"FR","type":"government","lineage":["https://openalex.org/I1294671590"]},{"id":"https://openalex.org/I4405277525","display_name":"Centre de Recherche en Informatique de Lens","ror":"https://ror.org/03072md54","country_code":"FR","type":"facility","lineage":["https://openalex.org/I1294671590","https://openalex.org/I4405277525","https://openalex.org/I44563897"]},{"id":"https://openalex.org/I44563897","display_name":"Universit\u00e9 d'Artois","ror":"https://ror.org/053x9s498","country_code":"FR","type":"education","lineage":["https://openalex.org/I44563897"]}],"countries":["FR"],"is_corresponding":false,"raw_author_name":"Haas, Nathana\u00ebl","raw_affiliation_strings":["Universit\u00e9 d'Artois , CNRS , CRIL UMR 8188 , Lens , France"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Universit\u00e9 d'Artois , CNRS , CRIL UMR 8188 , Lens , France","institution_ids":["https://openalex.org/I1294671590","https://openalex.org/I4405277525","https://openalex.org/I44563897"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5059764145","display_name":"Zied Bouraoui","orcid":"https://orcid.org/0000-0002-1662-4163","observed_orcids":["https://orcid.org/0000-0002-1662-4163"]},"institutions":[{"id":"https://openalex.org/I1294671590","display_name":"Centre National de la Recherche Scientifique","ror":"https://ror.org/02feahw73","country_code":"FR","type":"government","lineage":["https://openalex.org/I1294671590"]},{"id":"https://openalex.org/I4405277525","display_name":"Centre de Recherche en Informatique de Lens","ror":"https://ror.org/03072md54","country_code":"FR","type":"facility","lineage":["https://openalex.org/I1294671590","https://openalex.org/I4405277525","https://openalex.org/I44563897"]},{"id":"https://openalex.org/I44563897","display_name":"Universit\u00e9 d'Artois","ror":"https://ror.org/053x9s498","country_code":"FR","type":"education","lineage":["https://openalex.org/I44563897"]}],"countries":["FR"],"is_corresponding":false,"raw_author_name":"Bouraoui, Zied","raw_affiliation_strings":["Universit\u00e9 d'Artois , CNRS , CRIL UMR 8188 , Lens , France"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Universit\u00e9 d'Artois , CNRS , CRIL UMR 8188 , Lens , France","institution_ids":["https://openalex.org/I1294671590","https://openalex.org/I4405277525","https://openalex.org/I44563897"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":3,"corresponding_author_ids":[],"corresponding_institution_ids":[],"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":0.9878000020980835,"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":0.9878000020980835,"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/T10028","display_name":"Topic Modeling","score":0.0027000000700354576,"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/T10320","display_name":"Neural Networks and Applications","score":0.0017999999690800905,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/transformer-compression","display_name":"transformer compression","score":0.7210000157356262},{"id":"https://openalex.org/keywords/gauge-freedom","display_name":"gauge freedom","score":0.6639999747276306},{"id":"https://openalex.org/keywords/structural-complexity","display_name":"structural complexity","score":0.34299999475479126},{"id":"https://openalex.org/keywords/training-acceleration","display_name":"training acceleration","score":0.7789999842643738}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6351000070571899},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.5199000239372253},{"id":"https://openalex.org/C205711294","wikidata":"https://www.wikidata.org/wiki/Q176953","display_name":"Rendering (computer graphics)","level":2,"score":0.4876999855041504},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.48399999737739563},{"id":"https://openalex.org/C90119067","wikidata":"https://www.wikidata.org/wiki/Q43260","display_name":"Polynomial","level":2,"score":0.39340001344680786},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.3727000057697296},{"id":"https://openalex.org/C132459708","wikidata":"https://www.wikidata.org/wiki/Q744069","display_name":"Extrapolation","level":2,"score":0.3637999892234802},{"id":"https://openalex.org/C155281189","wikidata":"https://www.wikidata.org/wiki/Q3518150","display_name":"Tensor (intrinsic definition)","level":2,"score":0.3612000048160553},{"id":"https://openalex.org/C127162648","wikidata":"https://www.wikidata.org/wiki/Q16858953","display_name":"Channel (broadcasting)","level":2,"score":0.3605000078678131},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.35580000281333923},{"id":"https://openalex.org/C152124472","wikidata":"https://www.wikidata.org/wiki/Q1204361","display_name":"Redundancy (engineering)","level":2,"score":0.3531000018119812},{"id":"https://openalex.org/C200331156","wikidata":"https://www.wikidata.org/wiki/Q506041","display_name":"Jacobian matrix and determinant","level":2,"score":0.34790000319480896},{"id":"https://openalex.org/C2834757","wikidata":"https://www.wikidata.org/wiki/Q4925424","display_name":"Monotone polygon","level":2,"score":0.3402999937534332},{"id":"https://openalex.org/C175291020","wikidata":"https://www.wikidata.org/wiki/Q1156822","display_name":"Offset (computer science)","level":2,"score":0.33390000462532043},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.314300000667572},{"id":"https://openalex.org/C184720557","wikidata":"https://www.wikidata.org/wiki/Q7825049","display_name":"Topology (electrical circuits)","level":2,"score":0.30959999561309814},{"id":"https://openalex.org/C136886441","wikidata":"https://www.wikidata.org/wiki/Q926129","display_name":"Normalization (sociology)","level":2,"score":0.3003999888896942},{"id":"https://openalex.org/C180016635","wikidata":"https://www.wikidata.org/wiki/Q2712821","display_name":"Compression (physics)","level":2,"score":0.29499998688697815},{"id":"https://openalex.org/C184898388","wikidata":"https://www.wikidata.org/wiki/Q1435712","display_name":"Pairwise comparison","level":2,"score":0.29490000009536743},{"id":"https://openalex.org/C193415008","wikidata":"https://www.wikidata.org/wiki/Q639681","display_name":"Network architecture","level":2,"score":0.28630000352859497},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.28540000319480896},{"id":"https://openalex.org/C165838908","wikidata":"https://www.wikidata.org/wiki/Q736777","display_name":"Calibration","level":2,"score":0.2818000018596649},{"id":"https://openalex.org/C25797200","wikidata":"https://www.wikidata.org/wiki/Q828137","display_name":"Compression ratio","level":3,"score":0.28029999136924744},{"id":"https://openalex.org/C14036430","wikidata":"https://www.wikidata.org/wiki/Q3736076","display_name":"Function (biology)","level":2,"score":0.27619999647140503},{"id":"https://openalex.org/C70970002","wikidata":"https://www.wikidata.org/wiki/Q189434","display_name":"Multiplexer","level":3,"score":0.27619999647140503},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.26420000195503235},{"id":"https://openalex.org/C118505674","wikidata":"https://www.wikidata.org/wiki/Q42586063","display_name":"Encoder","level":2,"score":0.26420000195503235},{"id":"https://openalex.org/C108010975","wikidata":"https://www.wikidata.org/wiki/Q500094","display_name":"Pruning","level":2,"score":0.2603999972343445},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.25920000672340393},{"id":"https://openalex.org/C92757383","wikidata":"https://www.wikidata.org/wiki/Q382497","display_name":"Affine transformation","level":2,"score":0.2572000026702881},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.2522999942302704},{"id":"https://openalex.org/C189430467","wikidata":"https://www.wikidata.org/wiki/Q7293293","display_name":"Ranking (information retrieval)","level":2,"score":0.25189998745918274},{"id":"https://openalex.org/C157202957","wikidata":"https://www.wikidata.org/wiki/Q1659609","display_name":"Image warping","level":2,"score":0.25119999051094055},{"id":"https://openalex.org/C509619924","wikidata":"https://www.wikidata.org/wiki/Q17130757","display_name":"Structural complexity","level":2,"score":0.25119999051094055}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2609.07264","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.07264","pdf_url":"https://arxiv.org/pdf/2609.07264","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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2609.07264","doi":"https://doi.org/10.48550/arxiv.2609.07264","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.07264","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:2609.07264","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.07264","pdf_url":"https://arxiv.org/pdf/2609.07264","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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[{"id":"https://openalex.org/G589531907","display_name":"Event-Centric Reasoning for Interpreting Everyday Narratives","funder_award_id":"ANR-22-CE23-0002","funder_id":"https://openalex.org/F4320320883","funder_display_name":"Agence Nationale de la Recherche"},{"id":"https://openalex.org/G7712972687","display_name":"Mastering AI Applications","funder_award_id":"ANR-22-EXES-0009","funder_id":"https://openalex.org/F4320320883","funder_display_name":"Agence Nationale de la Recherche"}],"funders":[{"id":"https://openalex.org/F4320320883","display_name":"Agence Nationale de la Recherche","ror":"https://ror.org/00rbzpz17"},{"id":"https://openalex.org/F4320326256","display_name":"Grand \u00c9quipement National De Calcul Intensif","ror":"https://ror.org/0010d1q40"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7212059249.pdf","grobid_xml":"https://content.openalex.org/works/W7212059249.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7140193867","https://openalex.org/W7164858881","https://openalex.org/W7161122883","https://openalex.org/W7124446557","https://openalex.org/W7127352613","https://openalex.org/W7155386108","https://openalex.org/W4395044036","https://openalex.org/W7220366932","https://openalex.org/W4406178928","https://openalex.org/W7149340637"],"abstract_inverted_index":{"Inference":[0],"energy":[1],"per":[2,58,138],"token":[3],"drives":[4],"the":[5,77,84,117,144,157,162,187,218,224],"cost":[6],"and":[7,26,76,204],"carbon":[8],"footprint":[9],"of":[10,86,114,240],"deployed":[11],"transformers.":[12],"It":[13],"is":[14,93,236],"dominated":[15],"by":[16,96],"dense":[17,35],"matrix":[18],"products":[19],"that":[20,109,133,151,233],"incur":[21],"fused":[22],"multiply-accumulate":[23],"(FMA)":[24],"operations":[25],"memory":[27],"traffic.":[28],"To":[29],"reduce":[30],"these":[31,230],"computations":[32],"while":[33,73],"retaining":[34],"tensors":[36],"for":[37],"high":[38],"GPU":[39],"throughput,":[40],"we":[41],"develop":[42],"a":[43,111,196,208,244],"methodology":[44],"from":[45,172],"first":[46],"principles":[47],"to":[48,54,67,69,120,129,160,174,238],"adapt":[49],"structural":[50,149],"complexity":[51],"during":[52],"training":[53,215],"maximize":[55],"inference":[56,136],"utility":[57,137,226],"unit":[59,139],"compute.":[60,140],"Channel":[61],"penalties":[62,108],"drive":[63],"entire":[64],"tensor":[65],"slices":[66,132],"zero":[68],"enable":[70],"physical":[71],"removal":[72],"preserving":[74],"density":[75],"network":[78,118,145],"function.":[79],"The":[80],"natural":[81],"approach,":[82],"penalizing":[83],"norm":[85],"operator":[87],"components":[88],"acting":[89],"through":[90],"each":[91],"channel,":[92],"provably":[94],"destabilized":[95],"gauge":[97,121],"freedom.":[98],"We":[99],"resolve":[100],"this":[101],"pathology":[102],"with":[103,177,223],"GaugeLasso:":[104],"additive":[105],"symmetric":[106],"group-lasso":[107],"recover":[110],"monotone":[112],"function":[113],"product-norms":[115],"when":[116],"converges":[119],"balance.":[122],"Our":[123],"equilibrium":[124],"analysis":[125],"enables":[126],"per-channel":[127],"calibration":[128],"correctly":[130],"suppress":[131],"under-perform":[134],"in":[135],"Under":[141],"adaptive":[142],"pressure,":[143],"reorganizes":[146],"into":[147],"depth-dependent":[148],"profiles":[150],"can":[152],"be":[153],"far":[154],"smaller":[155],"than":[156],"architecture":[158,246],"required":[159],"learn":[161],"task.":[163,260],"On":[164,180,193],"polynomial":[165],"long":[166],"division":[167],"over":[168],"$\\mathbb{F}_{31}$,":[169],"compute":[170],"compresses":[171],"148":[173],"255":[175],"times":[176],"perfect":[178],"accuracy.":[179],"character-level":[181],"language":[182],"modeling,":[183],"compressed":[184],"models":[185],"outperform":[186],"hand-designed":[188],"baseline":[189,248],"at":[190],"equal":[191],"FMA.":[192],"masked":[194],"autoencoding,":[195],"compression":[197],"trial":[198],"exposes":[199],"which":[200,205],"axes":[201],"were":[202],"over-provisioned":[203],"saturated,":[206],"guiding":[207],"better":[209],"second":[210],"design.":[211],"Compaction":[212],"also":[213],"accelerates":[214],"monotonically":[216],"as":[217],"model":[219],"progresses.":[220],"Post-hoc":[221],"pruning":[222],"same":[225],"ranking":[227],"cannot":[228],"reach":[229],"structures,":[231],"showing":[232],"sustained":[234],"pressure":[235],"central":[237],"discovery":[239],"efficient":[241],"models.":[242],"Retraining":[243],"discovered":[245],"recovers":[247],"quality":[249],"on":[250,256],"our":[251,257],"statistical":[252],"tasks,":[253],"but":[254],"fails":[255],"exact":[258],"algorithmic":[259]},"counts_by_year":[],"updated_date":"2026-11-05T12:25:35.769452","created_date":"2026-09-10T00:00:00"}