{"id":"https://openalex.org/W7201865495","doi":"https://doi.org/10.48550/arxiv.2608.05926","title":"BALANCE: Hybrid Autoregressive-Speculative LLM Inference at the Network Edge","display_name":"BALANCE: Hybrid Autoregressive-Speculative LLM Inference at the Network Edge","publication_year":2026,"publication_date":"2026-08-06","ids":{"openalex":"https://openalex.org/W7201865495","doi":"https://doi.org/10.48550/arxiv.2608.05926"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2608.05926","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.05926","pdf_url":"https://arxiv.org/pdf/2608.05926","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","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2608.05926","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5023676698","display_name":"Guanqiao Qu","orcid":"https://orcid.org/0000-0002-5291-3317","observed_orcids":["https://orcid.org/0000-0002-5291-3317"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Qu, Guanqiao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5145994677","display_name":"Shuo Chen","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chen, Shuo","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5121348981","display_name":"Qian Chen","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chen, Qian","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5020917506","display_name":"Kin Kwong Leung","orcid":"https://orcid.org/0000-0002-3860-6257","observed_orcids":["https://orcid.org/0000-0002-3860-6257"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Leung, Kin K.","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5125726055","display_name":"Xianhao Chen","orcid":"https://orcid.org/0000-0002-4295-940X","observed_orcids":["https://orcid.org/0000-0002-4295-940X"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chen, Xianhao","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":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/T10273","display_name":"IoT and Edge/Fog Computing","score":0.9854000210762024,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"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/T10273","display_name":"IoT and Edge/Fog Computing","score":0.9854000210762024,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"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/T10181","display_name":"Natural Language Processing Techniques","score":0.0019000000320374966,"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/T14098","display_name":"Energy Efficiency in Computing","score":0.0010000000474974513,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"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/large-language-models","display_name":"large language models","score":0.7929999828338623},{"id":"https://openalex.org/keywords/speculative-decoding","display_name":"speculative decoding","score":0.8009999990463257},{"id":"https://openalex.org/keywords/autoregressive-decoding","display_name":"autoregressive decoding","score":1.0},{"id":"https://openalex.org/keywords/hybrid-inference","display_name":"hybrid inference","score":0.7110000252723694},{"id":"https://openalex.org/keywords/user-admission-control","display_name":"user admission control","score":0.8029999732971191},{"id":"https://openalex.org/keywords/computational-resource-allocation","display_name":"computing resource allocation","score":0.8140000104904175},{"id":"https://openalex.org/keywords/approximation-algorithms","display_name":"approximation algorithms","score":0.7319999933242798},{"id":"https://openalex.org/keywords/cellular-networks","display_name":"cellular networks","score":0.6850000023841858},{"id":"https://openalex.org/keywords/edge-computing","display_name":"edge computing","score":0.5400000214576721}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.8457000255584717},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.7682999968528748},{"id":"https://openalex.org/C206729178","wikidata":"https://www.wikidata.org/wiki/Q2271896","display_name":"Scheduling (production processes)","level":2,"score":0.5461999773979187},{"id":"https://openalex.org/C2778456923","wikidata":"https://www.wikidata.org/wiki/Q5337692","display_name":"Edge computing","level":3,"score":0.5382999777793884},{"id":"https://openalex.org/C162307627","wikidata":"https://www.wikidata.org/wiki/Q204833","display_name":"Enhanced Data Rates for GSM Evolution","level":2,"score":0.5051000118255615},{"id":"https://openalex.org/C82876162","wikidata":"https://www.wikidata.org/wiki/Q17096504","display_name":"Latency (audio)","level":2,"score":0.4724000096321106},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.4661000072956085},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.4456999897956848},{"id":"https://openalex.org/C2777472644","wikidata":"https://www.wikidata.org/wiki/Q16968992","display_name":"Approximate inference","level":3,"score":0.40139999985694885},{"id":"https://openalex.org/C555944384","wikidata":"https://www.wikidata.org/wiki/Q249","display_name":"Wireless","level":2,"score":0.3896999955177307},{"id":"https://openalex.org/C138236772","wikidata":"https://www.wikidata.org/wiki/Q25098575","display_name":"Edge device","level":3,"score":0.3864000141620636},{"id":"https://openalex.org/C120314980","wikidata":"https://www.wikidata.org/wiki/Q180634","display_name":"Distributed computing","level":1,"score":0.37049999833106995},{"id":"https://openalex.org/C157764524","wikidata":"https://www.wikidata.org/wiki/Q1383412","display_name":"Throughput","level":3,"score":0.3472000062465668},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.33550000190734863},{"id":"https://openalex.org/C108037233","wikidata":"https://www.wikidata.org/wiki/Q11375","display_name":"Wireless network","level":3,"score":0.32440000772476196},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.3100999891757965},{"id":"https://openalex.org/C137836250","wikidata":"https://www.wikidata.org/wiki/Q984063","display_name":"Optimization problem","level":2,"score":0.28700000047683716},{"id":"https://openalex.org/C175154964","wikidata":"https://www.wikidata.org/wiki/Q380077","display_name":"Task analysis","level":3,"score":0.28690001368522644},{"id":"https://openalex.org/C2776330181","wikidata":"https://www.wikidata.org/wiki/Q18358244","display_name":"Maximization","level":2,"score":0.27140000462532043},{"id":"https://openalex.org/C31258907","wikidata":"https://www.wikidata.org/wiki/Q1301371","display_name":"Computer network","level":1,"score":0.26969999074935913},{"id":"https://openalex.org/C2776061582","wikidata":"https://www.wikidata.org/wiki/Q25325231","display_name":"Mobile edge computing","level":3,"score":0.2655999958515167},{"id":"https://openalex.org/C93996380","wikidata":"https://www.wikidata.org/wiki/Q44127","display_name":"Server","level":2,"score":0.2628999948501587},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.25459998846054077}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2608.05926","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.05926","pdf_url":"https://arxiv.org/pdf/2608.05926","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.2608.05926","doi":"https://doi.org/10.48550/arxiv.2608.05926","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2608.05926","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2608.05926","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.05926","pdf_url":"https://arxiv.org/pdf/2608.05926","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":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7201865495.pdf","grobid_xml":"https://content.openalex.org/works/W7201865495.grobid-xml"},"referenced_works_count":4,"referenced_works":["https://openalex.org/W4407358710","https://openalex.org/W4416435460","https://openalex.org/W7155554337","https://openalex.org/W7168147450"],"related_works":["https://openalex.org/W7091187908","https://openalex.org/W7104018857","https://openalex.org/W7155186996","https://openalex.org/W4400277194","https://openalex.org/W7128723494","https://openalex.org/W7128997886","https://openalex.org/W4417282938","https://openalex.org/W4404562740","https://openalex.org/W4414110614","https://openalex.org/W7155554337"],"abstract_inverted_index":{"Edge":[0],"inference":[1,12,19,40,90],"is":[2,165],"a":[3,43,87,136,169,182,186],"promising":[4],"paradigm":[5],"to":[6,48,62,115,141],"provide":[7],"large":[8],"language":[9,45],"model":[10,46],"(LLM)":[11],"services":[13],"in":[14,33],"next-generation":[15],"mobile":[16],"networks.":[17],"LLM":[18,54,95],"mainly":[20],"relies":[21],"on":[22],"two":[23,124,178],"approaches:":[24],"Autoregressive":[25],"decoding":[26,37],"(AD)":[27],"generates":[28],"output":[29],"tokens":[30,52],"sequentially,":[31],"resulting":[32],"long":[34],"latency;":[35],"Speculative":[36],"(SD)":[38],"accelerates":[39],"by":[41],"using":[42],"small":[44],"(SLM)":[47],"generate":[49],"multiple":[50],"draft":[51],"for":[53,93],"verification,":[55],"but":[56],"incurs":[57],"extra":[58],"memory":[59,160],"costs.":[60],"Due":[61],"this":[63,83],"latency-memory":[64],"tradeoff,":[65],"neither":[66],"approach":[67],"alone":[68],"can":[69],"efficiently":[70],"serve":[71],"users":[72],"with":[73,185],"heterogeneous":[74],"demands":[75],"under":[76,154],"limited":[77],"edge":[78,94,100],"computing":[79,147],"resources.":[80],"To":[81,127],"address":[82],"challenge,":[84],"we":[85,134,167],"propose":[86],"hybrid":[88],"autoregressive-speculative":[89],"(BALANCE)":[91],"framework":[92],"inference.":[96],"In":[97],"BALANCE,":[98],"an":[99,104,107],"server":[101,159],"hosts":[102],"both":[103],"SLM":[105],"and":[106,121,146,152,158,180,198,200],"LLM,":[108],"admits":[109],"users,":[110,133],"assigns":[111],"each":[112],"admitted":[113],"user":[114,144,155],"the":[116,123,129,163,174],"AD":[117,151,197],"or":[118],"SD":[119,153,199],"mode,":[120],"performs":[122],"modes":[125],"simultaneously.":[126],"maximize":[128],"number":[130],"of":[131],"served":[132],"formulate":[135],"task":[137,203],"throughput":[138],"maximization":[139],"problem":[140,164,176],"jointly":[142],"determine":[143],"admission":[145],"resource":[148],"allocation":[149],"between":[150],"latency":[156],"requirements":[157],"constraints.":[161],"Since":[162],"NP-hard,":[166],"develop":[168],"polynomial-time":[170],"algorithm":[171],"that":[172,192],"transforms":[173],"original":[175],"into":[177],"sub-problems":[179],"obtains":[181],"sub-optimal":[183],"solution":[184],"constant":[187],"approximation":[188],"guarantee.":[189],"Experiments":[190],"demonstrate":[191],"BALANCE":[193],"consistently":[194],"outperforms":[195],"conventional":[196],"significantly":[201],"improves":[202],"throughput.":[204]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-08-08T00:00:00"}