{"id":"https://openalex.org/W4414857751","doi":"https://doi.org/10.48550/arxiv.2505.24544","title":"Cross-Attention Speculative Decoding","display_name":"Cross-Attention Speculative Decoding","publication_year":2025,"publication_date":"2025-05-30","ids":{"openalex":"https://openalex.org/W4414857751","doi":"https://doi.org/10.48550/arxiv.2505.24544"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2505.24544","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2505.24544","pdf_url":"https://arxiv.org/pdf/2505.24544","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/2505.24544","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Zhong, Wei","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhong, Wei","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5024212405","display_name":"Manasa Bharadwaj","orcid":"https://orcid.org/0009-0002-0128-4706","observed_orcids":["https://orcid.org/0009-0002-0128-4706"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Bharadwaj, Manasa","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100651550","display_name":"Yixiao Wang","orcid":"https://orcid.org/0000-0003-2664-3605","observed_orcids":["https://orcid.org/0000-0003-2664-3605"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Yixiao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5119862709","display_name":"Yipeng Ji","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ji, Yipeng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5049821500","display_name":"C.M. Lee","orcid":"https://orcid.org/0000-0001-9329-7365","observed_orcids":["https://orcid.org/0000-0001-9329-7365"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lee, Chul","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/T10181","display_name":"Natural Language Processing Techniques","score":0.8008000254631042,"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"}},"topics":[{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.8008000254631042,"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/T10028","display_name":"Topic Modeling","score":0.1517000049352646,"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":0.01979999989271164,"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/speculative-decoding","display_name":"speculative decoding","score":0.9950000047683716},{"id":"https://openalex.org/keywords/cross-attention","display_name":"cross-attention","score":0.9049999713897705},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.8080000281333923},{"id":"https://openalex.org/keywords/inference-acceleration","display_name":"inference acceleration","score":0.7509999871253967},{"id":"https://openalex.org/keywords/transformer-decoder","display_name":"transformer decoder","score":0.42399999499320984},{"id":"https://openalex.org/keywords/training-stability","display_name":"training stability","score":0.6370000243186951},{"id":"https://openalex.org/keywords/memory-usage","display_name":"memory usage","score":0.5099999904632568},{"id":"https://openalex.org/keywords/beagle","display_name":"beagle dogs","score":0.628000020980835}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7609000205993652},{"id":"https://openalex.org/C70437156","wikidata":"https://www.wikidata.org/wiki/Q7228652","display_name":"Pooling","level":2,"score":0.6725000143051147},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.633400022983551},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5440999865531921},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.5293999910354614},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4359999895095825},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.40220001339912415},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.36010000109672546},{"id":"https://openalex.org/C206729178","wikidata":"https://www.wikidata.org/wiki/Q2271896","display_name":"Scheduling (production processes)","level":2,"score":0.34619998931884766},{"id":"https://openalex.org/C154690210","wikidata":"https://www.wikidata.org/wiki/Q1668499","display_name":"Rewriting","level":2,"score":0.32820001244544983},{"id":"https://openalex.org/C112972136","wikidata":"https://www.wikidata.org/wiki/Q7595718","display_name":"Stability (learning theory)","level":2,"score":0.32589998841285706},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.31949999928474426},{"id":"https://openalex.org/C45493050","wikidata":"https://www.wikidata.org/wiki/Q7884934","display_name":"Unified Model","level":2,"score":0.28619998693466187},{"id":"https://openalex.org/C2780069185","wikidata":"https://www.wikidata.org/wiki/Q7977945","display_name":"Equivalence (formal languages)","level":2,"score":0.28459998965263367},{"id":"https://openalex.org/C2777303404","wikidata":"https://www.wikidata.org/wiki/Q759757","display_name":"Convergence (economics)","level":2,"score":0.27140000462532043},{"id":"https://openalex.org/C12186640","wikidata":"https://www.wikidata.org/wiki/Q6815743","display_name":"Memory model","level":3,"score":0.2694999873638153},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.266400009393692},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.2524999976158142}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2505.24544","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2505.24544","pdf_url":"https://arxiv.org/pdf/2505.24544","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.2505.24544","doi":"https://doi.org/10.48550/arxiv.2505.24544","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2505.24544","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":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2505.24544","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2505.24544","pdf_url":"https://arxiv.org/pdf/2505.24544","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":[],"sustainable_development_goals_aurora":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Speculative":[0],"decoding":[1],"(SD)":[2],"is":[3],"a":[4,121,156],"widely":[5],"adopted":[6],"approach":[7],"for":[8,89,159],"accelerating":[9],"inference":[10,147],"in":[11,131,161],"large":[12],"language":[13],"models":[14,22,83],"(LLMs),":[15],"particularly":[16],"when":[17],"the":[18,64,87,95],"draft":[19],"and":[20,52,100,128,140,149],"target":[21],"are":[23],"well":[24],"aligned.":[25],"However,":[26],"state-of-the-art":[27],"SD":[28,72,82],"methods":[29],"typically":[30],"rely":[31],"on":[32,77],"tightly":[33],"coupled,":[34],"self-attention-based":[35],"Transformer":[36,70],"decoders,":[37],"often":[38],"augmented":[39],"with":[40,79],"auxiliary":[41,92],"pooling":[42,90],"or":[43,91],"fusion":[44],"layers.":[45],"This":[46],"coupling":[47],"makes":[48],"them":[49],"increasingly":[50],"complex":[51],"harder":[53],"to":[54,66],"generalize":[55],"across":[56,137],"different":[57],"models.":[58],"We":[59],"present":[60],"Budget":[61],"EAGLE":[62],"(Beagle),":[63],"first,":[65],"our":[67],"knowledge,":[68],"cross-attention-based":[69],"decoder":[71],"model":[73],"that":[74,124,143],"achieves":[75,125,145],"performance":[76],"par":[78],"leading":[80],"self-attention":[81],"(EAGLE-v2)":[84],"while":[85],"eliminating":[86],"need":[88],"components,":[93],"simplifying":[94],"architecture,":[96,115],"improving":[97],"training":[98,111,126,151],"efficiency,":[99],"maintaining":[101],"stable":[102],"memory":[103],"usage":[104],"during":[105],"training-time":[106],"simulation.":[107],"To":[108],"enable":[109],"effective":[110],"of":[112],"this":[113],"novel":[114],"we":[116],"propose":[117],"Two-Stage":[118],"Block-Attention":[119],"Training,":[120],"new":[122],"method":[123],"stability":[127],"convergence":[129],"efficiency":[130,152],"block-level":[132],"attention":[133],"scenarios.":[134],"Extensive":[135],"experiments":[136],"multiple":[138],"LLMs":[139],"datasets":[141],"show":[142],"Beagle":[144],"competitive":[146],"speedups":[148],"higher":[150],"than":[153],"EAGLE-v2,":[154],"offering":[155],"strong":[157],"alternative":[158],"architectures":[160],"speculative":[162],"decoding.":[163]},"counts_by_year":[],"updated_date":"2026-09-29T14:09:16.559479","created_date":"2025-10-10T00:00:00"}