{"id":"https://openalex.org/W7143293890","doi":"https://doi.org/10.48550/arxiv.2603.26556","title":"When Perplexity Lies: Generation-Focused Distillation of Hybrid Sequence Models","display_name":"When Perplexity Lies: Generation-Focused Distillation of Hybrid Sequence Models","publication_year":2026,"publication_date":"2026-03-27","ids":{"openalex":"https://openalex.org/W7143293890","doi":"https://doi.org/10.48550/arxiv.2603.26556"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2603.26556","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.26556","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,"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.2603.26556","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5131001817","display_name":"Juan Gabriel Kostelec","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Kostelec, Juan Gabriel","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5141783261","display_name":"Qinghai Guo","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Guo, Qinghai","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/T11948","display_name":"Machine Learning in Materials Science","score":0.2856999933719635,"subfield":{"id":"https://openalex.org/subfields/2505","display_name":"Materials Chemistry"},"field":{"id":"https://openalex.org/fields/25","display_name":"Materials Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T11948","display_name":"Machine Learning in Materials Science","score":0.2856999933719635,"subfield":{"id":"https://openalex.org/subfields/2505","display_name":"Materials Chemistry"},"field":{"id":"https://openalex.org/fields/25","display_name":"Materials Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.14079999923706055,"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/T12535","display_name":"Machine Learning and Data Classification","score":0.11089999973773956,"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/distillation","display_name":"Distillation","score":0.7016000151634216},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.546999990940094},{"id":"https://openalex.org/keywords/transformer","display_name":"Transformer","score":0.4837999939918518},{"id":"https://openalex.org/keywords/autoregressive-model","display_name":"Autoregressive model","score":0.41659998893737793},{"id":"https://openalex.org/keywords/perplexity","display_name":"Perplexity","score":0.415800005197525},{"id":"https://openalex.org/keywords/classifier","display_name":"Classifier (UML)","score":0.38370001316070557}],"concepts":[{"id":"https://openalex.org/C204030448","wikidata":"https://www.wikidata.org/wiki/Q101017","display_name":"Distillation","level":2,"score":0.7016000151634216},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6960999965667725},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.546999990940094},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.5174000263214111},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5083000063896179},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.4837999939918518},{"id":"https://openalex.org/C159877910","wikidata":"https://www.wikidata.org/wiki/Q2202883","display_name":"Autoregressive model","level":2,"score":0.41659998893737793},{"id":"https://openalex.org/C100279451","wikidata":"https://www.wikidata.org/wiki/Q372193","display_name":"Perplexity","level":3,"score":0.415800005197525},{"id":"https://openalex.org/C95623464","wikidata":"https://www.wikidata.org/wiki/Q1096149","display_name":"Classifier (UML)","level":2,"score":0.38370001316070557},{"id":"https://openalex.org/C189430467","wikidata":"https://www.wikidata.org/wiki/Q7293293","display_name":"Ranking (information retrieval)","level":2,"score":0.3822999894618988},{"id":"https://openalex.org/C2778112365","wikidata":"https://www.wikidata.org/wiki/Q3511065","display_name":"Sequence (biology)","level":2,"score":0.31049999594688416},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.2881999909877777},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.28600001335144043},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.27230000495910645},{"id":"https://openalex.org/C123657996","wikidata":"https://www.wikidata.org/wiki/Q12271","display_name":"Architecture","level":2,"score":0.2676999866962433}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2603.26556","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.26556","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,"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.2603.26556","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.26556","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,"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":[{"display_name":"Quality Education","id":"https://metadata.un.org/sdg/4","score":0.6818745732307434}],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Converting":[0],"a":[1,5,13,75,108,115,130,224],"pretrained":[2,116],"Transformer":[3,117],"into":[4,118],"more":[6],"efficient":[7,120],"hybrid":[8],"model":[9,65,78,226],"through":[10],"distillation":[11,38,42,221],"offers":[12],"promising":[14],"approach":[15],"to":[16,84,245],"reducing":[17,239],"inference":[18],"costs.":[19],"However,":[20],"achieving":[21],"high-quality":[22],"generation":[23,217],"in":[24,64,179],"distilled":[25,77],"models":[26],"requires":[27],"careful":[28],"joint":[29],"design":[30,139,186],"of":[31,185,232],"both":[32,159],"the":[33,37,171,183,198,213,228],"student":[34],"architecture":[35,152],"and":[36,151,154,161,175,177,206,247],"process.":[39],"Many":[40],"prior":[41],"works":[43],"evaluate":[44,155],"downstream":[45],"multiple-choice":[46],"benchmarks":[47,237],"by":[48,93,243,250],"ranking":[49,184],"candidate":[50],"answers":[51,100],"with":[52,106],"log-likelihood":[53,89,160],"rather":[54],"than":[55],"requiring":[56],"autoregressive":[57],"generation,":[58],"which":[59],"can":[60,178],"obscure":[61],"important":[62],"differences":[63],"quality.":[66,218],"For":[67],"example,":[68],"on":[69,133,216,235],"overlapping":[70],"benchmarks,":[71],"we":[72,111,135,200],"show":[73],"that":[74,79,166],"7B":[76],"nearly":[80],"matches":[81],"its":[82],"teacher":[83,174,233],"within":[85],"0.2":[86],"pp":[87,95],"under":[88,158],"scoring":[90],"falls":[91],"behind":[92],"20.8":[94],"when":[96],"it":[97,128],"must":[98],"generate":[99],"autoregressively.":[101],"We":[102,164],"investigate":[103],"this":[104],"phenomenon":[105],"GenDistill,":[107],"multi-stage":[109],"pipeline":[110],"designed":[112],"for":[113],"distilling":[114],"an":[119],"Hybrid":[121],"Kimi":[122],"Delta":[123],"Attention":[124],"(Hybrid-KDA)":[125],"student.":[126],"Using":[127],"as":[129,227],"controlled":[131],"testbed":[132],"Qwen3-0.6B,":[134],"systematically":[136],"ablate":[137],"six":[138],"axes":[140],"(training":[141],"objective,":[142],"loss":[143],"masking,":[144,205],"training":[145],"duration,":[146],"dataset":[147,202],"selection,":[148,203],"parameter":[149],"freezing,":[150],"choice)":[153],"every":[156],"choice":[157],"generation-based":[162],"protocols.":[163],"find":[165],"log-likelihood-based":[167],"evaluation":[168,193],"consistently":[169],"underestimates":[170],"gap":[172],"between":[173],"student,":[176,229],"some":[180],"cases":[181],"reverse":[182],"choices,":[187],"so":[188],"conclusions":[189],"drawn":[190],"from":[191],"perplexity-only":[192],"may":[194],"be":[195],"misleading.":[196],"Among":[197],"factors":[199],"study,":[201],"completion-only":[204],"freezing":[207],"attention":[208],"layers":[209],"during":[210],"post-training":[211],"have":[212],"largest":[214],"impact":[215],"Our":[219],"best":[220],"recipe,":[222],"using":[223],"Hybrid-KDA":[225],"retains":[230],"86-90%":[231],"accuracy":[234],"knowledge":[236],"while":[238],"KV":[240],"cache":[241],"memory":[242],"up":[244],"75%":[246],"improving":[248],"time-to-first-token":[249],"2-4x":[251],"at":[252],"128K-token":[253],"contexts.":[254]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-03-31T00:00:00"}