{"id":"https://openalex.org/W7135220973","doi":"https://doi.org/10.48550/arxiv.2603.12038","title":"Slow-Fast Inference: Training-Free Inference Acceleration via Within-Sentence Support Stability","display_name":"Slow-Fast Inference: Training-Free Inference Acceleration via Within-Sentence Support Stability","publication_year":2026,"publication_date":"2026-03-12","ids":{"openalex":"https://openalex.org/W7135220973","doi":"https://doi.org/10.48550/arxiv.2603.12038"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2603.12038","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.12038","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.2603.12038","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5129048955","display_name":"Xingyu Xie","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xie, Xingyu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5129069444","display_name":"Zhaochen Yu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yu, Zhaochen","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5129025884","display_name":"Yue Liao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liao, Yue","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5129016140","display_name":"Tao Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Tao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Toh, Kim-Chuan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Toh, Kim-Chuan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5128956456","display_name":"Shuicheng Yan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yan, Shuicheng","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/T10028","display_name":"Topic Modeling","score":0.4837999939918518,"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/T10028","display_name":"Topic Modeling","score":0.4837999939918518,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.09719999879598618,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.057999998331069946,"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/decoding-methods","display_name":"Decoding methods","score":0.7731000185012817},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.6847000122070312},{"id":"https://openalex.org/keywords/context","display_name":"Context (archaeology)","score":0.6477000117301941},{"id":"https://openalex.org/keywords/autoregressive-model","display_name":"Autoregressive model","score":0.6148999929428101},{"id":"https://openalex.org/keywords/reuse","display_name":"Reuse","score":0.5511000156402588},{"id":"https://openalex.org/keywords/encoding","display_name":"Encoding (memory)","score":0.5101000070571899},{"id":"https://openalex.org/keywords/path","display_name":"Path (computing)","score":0.4855000078678131}],"concepts":[{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.7731000185012817},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6909000277519226},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6847000122070312},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.6477000117301941},{"id":"https://openalex.org/C159877910","wikidata":"https://www.wikidata.org/wiki/Q2202883","display_name":"Autoregressive model","level":2,"score":0.6148999929428101},{"id":"https://openalex.org/C206588197","wikidata":"https://www.wikidata.org/wiki/Q846574","display_name":"Reuse","level":2,"score":0.5511000156402588},{"id":"https://openalex.org/C125411270","wikidata":"https://www.wikidata.org/wiki/Q18653","display_name":"Encoding (memory)","level":2,"score":0.5101000070571899},{"id":"https://openalex.org/C2777735758","wikidata":"https://www.wikidata.org/wiki/Q817765","display_name":"Path (computing)","level":2,"score":0.4855000078678131},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.4383000135421753},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.421999990940094},{"id":"https://openalex.org/C112972136","wikidata":"https://www.wikidata.org/wiki/Q7595718","display_name":"Stability (learning theory)","level":2,"score":0.4162999987602234},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.40689998865127563},{"id":"https://openalex.org/C193969084","wikidata":"https://www.wikidata.org/wiki/Q7452500","display_name":"Sequential decoding","level":4,"score":0.3700000047683716},{"id":"https://openalex.org/C157764524","wikidata":"https://www.wikidata.org/wiki/Q1383412","display_name":"Throughput","level":3,"score":0.3361999988555908},{"id":"https://openalex.org/C111335779","wikidata":"https://www.wikidata.org/wiki/Q3454686","display_name":"Reduction (mathematics)","level":2,"score":0.3260999917984009},{"id":"https://openalex.org/C12725497","wikidata":"https://www.wikidata.org/wiki/Q810247","display_name":"Baseline (sea)","level":2,"score":0.31839999556541443},{"id":"https://openalex.org/C2779530757","wikidata":"https://www.wikidata.org/wiki/Q1207505","display_name":"Quality (philosophy)","level":2,"score":0.2833000123500824},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.2687000036239624},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.2556000053882599}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2603.12038","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.12038","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.2603.12038","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.12038","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":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Long-context":[0],"autoregressive":[1,155],"decoding":[2,7,53,117],"remains":[3,39],"expensive":[4],"because":[5],"each":[6],"step":[8],"must":[9],"repeatedly":[10],"process":[11],"a":[12,17,23,29,51,71,146],"growing":[13],"history.":[14],"We":[15],"observe":[16],"consistent":[18],"pattern":[19],"during":[20],"decoding:":[21],"within":[22,28],"sentence,":[24],"and":[25,63,94,131,138,161],"more":[26],"generally":[27,120],"short":[30],"semantically":[31],"coherent":[32],"span,":[33],"the":[34,88,91,96,100,108,126],"dominant":[35],"attention":[36],"support":[37],"often":[38],"largely":[40],"stable.":[41],"Motivated":[42],"by":[43],"this":[44],"observation,":[45],"we":[46],"propose":[47],"Slow-Fast":[48],"Inference":[49],"(SFI),":[50],"training-free":[52,137],"framework":[54],"that":[55],"decouples":[56],"generation":[57],"into":[58],"frequent":[59],"low-cost":[60],"fast":[61,105],"steps":[62,69,79],"occasional":[64],"dense-attention":[65],"slow":[66,86],"steps.":[67,106],"Fast":[68],"reuse":[70],"compact":[72],"sparse":[73],"memory":[74,102],"for":[75,103,153],"efficient":[76],"decoding.":[77],"Slow":[78],"are":[80],"triggered":[81],"near":[82],"semantic":[83],"boundaries.":[84],"At":[85],"steps,":[87],"model":[89],"revisits":[90],"broader":[92],"context":[93,110],"uses":[95],"Selector":[97],"to":[98,141,149],"refresh":[99],"selected":[101],"subsequent":[104],"Across":[107],"evaluated":[109],"lengths,":[111],"SFI":[112,135],"delivers":[113],"approximately":[114],"$1.6\\times$--$14.4\\times$":[115],"higher":[116],"throughput":[118],"while":[119],"maintaining":[121],"quality":[122],"on":[123],"par":[124],"with":[125],"full-KV":[127],"baseline":[128],"across":[129],"long-context":[130],"long-CoT":[132],"settings.":[133],"Because":[134],"is":[136],"applies":[139],"directly":[140],"existing":[142],"checkpoints,":[143],"it":[144],"offers":[145],"practical":[147],"path":[148],"reducing":[150],"inference":[151],"cost":[152],"contemporary":[154],"reasoning":[156],"models":[157],"in":[158],"long-context,":[159],"long-horizon,":[160],"agentic":[162],"workloads.":[163]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-03-14T00:00:00"}