{"id":"https://openalex.org/W7160940084","doi":"https://doi.org/10.48550/arxiv.2605.09932","title":"FocuSFT: Bilevel Optimization for Dilution-Aware Long-Context Fine-Tuning","display_name":"FocuSFT: Bilevel Optimization for Dilution-Aware Long-Context Fine-Tuning","publication_year":2026,"publication_date":"2026-05-11","ids":{"openalex":"https://openalex.org/W7160940084","doi":"https://doi.org/10.48550/arxiv.2605.09932"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.09932","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.09932","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.2605.09932","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5135943317","display_name":"Zehua Pei","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Pei, Zehua","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5135960920","display_name":"Hui-Ling Zhen","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhen, Hui-Ling","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5135967972","display_name":"Xianzhi Yu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yu, Xianzhi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5135927502","display_name":"Sinno Jialin Pan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Pan, Sinno Jialin","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5135996361","display_name":"Mingxuan Yuan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yuan, Mingxuan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5135942700","display_name":"Bei Yu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yu, Bei","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":0.27059999108314514,"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":0.27059999108314514,"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":0.23489999771118164,"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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.10080000013113022,"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/context","display_name":"Context (archaeology)","score":0.599399983882904},{"id":"https://openalex.org/keywords/process","display_name":"Process (computing)","score":0.5231999754905701},{"id":"https://openalex.org/keywords/context-model","display_name":"Context model","score":0.4406000077724457},{"id":"https://openalex.org/keywords/trace","display_name":"TRACE (psycholinguistics)","score":0.429500013589859},{"id":"https://openalex.org/keywords/parametric-statistics","display_name":"Parametric statistics","score":0.42879998683929443},{"id":"https://openalex.org/keywords/limiting","display_name":"Limiting","score":0.42570000886917114},{"id":"https://openalex.org/keywords/evasion","display_name":"Evasion (ethics)","score":0.39329999685287476},{"id":"https://openalex.org/keywords/partition","display_name":"Partition (number theory)","score":0.38019999861717224}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7372999787330627},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.599399983882904},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.5231999754905701},{"id":"https://openalex.org/C183322885","wikidata":"https://www.wikidata.org/wiki/Q17007702","display_name":"Context model","level":3,"score":0.4406000077724457},{"id":"https://openalex.org/C75291252","wikidata":"https://www.wikidata.org/wiki/Q1315756","display_name":"TRACE (psycholinguistics)","level":2,"score":0.429500013589859},{"id":"https://openalex.org/C117251300","wikidata":"https://www.wikidata.org/wiki/Q1849855","display_name":"Parametric statistics","level":2,"score":0.42879998683929443},{"id":"https://openalex.org/C188198153","wikidata":"https://www.wikidata.org/wiki/Q1613840","display_name":"Limiting","level":2,"score":0.42570000886917114},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.39719998836517334},{"id":"https://openalex.org/C2781251061","wikidata":"https://www.wikidata.org/wiki/Q5416089","display_name":"Evasion (ethics)","level":3,"score":0.39329999685287476},{"id":"https://openalex.org/C42812","wikidata":"https://www.wikidata.org/wiki/Q1082910","display_name":"Partition (number theory)","level":2,"score":0.38019999861717224},{"id":"https://openalex.org/C3309286","wikidata":"https://www.wikidata.org/wiki/Q4907693","display_name":"Bilevel optimization","level":3,"score":0.36239999532699585},{"id":"https://openalex.org/C2777402240","wikidata":"https://www.wikidata.org/wiki/Q6783436","display_name":"Masking (illustration)","level":2,"score":0.33239999413490295},{"id":"https://openalex.org/C137836250","wikidata":"https://www.wikidata.org/wiki/Q984063","display_name":"Optimization problem","level":2,"score":0.32510000467300415},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.31869998574256897},{"id":"https://openalex.org/C2776036281","wikidata":"https://www.wikidata.org/wiki/Q48769818","display_name":"Constraint (computer-aided design)","level":2,"score":0.29120001196861267},{"id":"https://openalex.org/C187191949","wikidata":"https://www.wikidata.org/wiki/Q1138496","display_name":"Profiling (computer programming)","level":2,"score":0.2777999937534332},{"id":"https://openalex.org/C76188268","wikidata":"https://www.wikidata.org/wiki/Q1783165","display_name":"Context effect","level":3,"score":0.2718000113964081},{"id":"https://openalex.org/C24574437","wikidata":"https://www.wikidata.org/wiki/Q7135228","display_name":"Parametric model","level":3,"score":0.27059999108314514},{"id":"https://openalex.org/C158154518","wikidata":"https://www.wikidata.org/wiki/Q7310970","display_name":"Relevance (law)","level":2,"score":0.26089999079704285},{"id":"https://openalex.org/C11671645","wikidata":"https://www.wikidata.org/wiki/Q5054567","display_name":"Causal model","level":2,"score":0.2540000081062317},{"id":"https://openalex.org/C163504300","wikidata":"https://www.wikidata.org/wiki/Q2364925","display_name":"Causal structure","level":2,"score":0.2515999972820282}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.09932","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.09932","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.2605.09932","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.09932","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":[{"id":"https://metadata.un.org/sdg/4","score":0.4434528648853302,"display_name":"Quality Education"}],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Large":[0],"language":[1],"models":[2],"can":[3],"now":[4],"process":[5],"increasingly":[6],"long":[7,18,37],"inputs,":[8],"yet":[9],"their":[10],"ability":[11,82],"to":[12,26,47,53,83,113,156,170,184],"effectively":[13],"use":[14],"information":[15],"spread":[16],"across":[17,172],"contexts":[19],"remains":[20],"limited.":[21],"We":[22,88],"trace":[23],"this":[24,97,132],"gap":[25],"how":[27],"attention":[28,42,52,64,73,120,139,157,209],"budget":[29],"is":[30],"spent":[31],"during":[32,218],"supervised":[33],"fine-tuning":[34],"(SFT)":[35],"on":[36,109,121,131,176,189],"sequences:":[38],"positional":[39],"biases":[40],"and":[41,124,159,188,214],"sinks":[43,158],"cause":[44],"the":[45,72,76,80,110,125,150],"model":[46],"allocate":[48],"most":[49],"of":[50,68],"its":[51],"positionally":[54],"privileged":[55],"tokens":[56,70,142],"rather":[57],"than":[58],"semantically":[59],"relevant":[60,122],"content.":[61],"This":[62],"training-time":[63],"dilution":[65],"(the":[66],"starvation":[67],"content":[69],"in":[71,201],"distribution)":[74],"weakens":[75],"gradient":[77],"signal,":[78],"limiting":[79],"model's":[81],"learn":[84],"robust":[85],"long-context":[86],"capabilities.":[87],"introduce":[89],"FocuSFT,":[90],"a":[91,115,197],"bilevel":[92],"optimization":[93],"framework":[94],"that":[95,118,153,206],"addresses":[96],"problem":[98],"at":[99,186],"training":[100,111],"time.":[101],"An":[102],"inner":[103],"loop":[104,127],"adapts":[105],"lightweight":[106],"fast-weight":[107],"parameters":[108],"context":[112,141,174,216],"form":[114],"parametric":[116],"memory":[117],"concentrates":[119],"content,":[123],"outer":[126],"performs":[128],"SFT":[129],"conditioned":[130],"sharpened":[133],"representation.":[134],"Both":[135],"loops":[136],"apply":[137],"bidirectional":[138],"over":[140],"while":[143],"preserving":[144],"causal":[145,151],"masking":[146],"for":[147],"responses,":[148],"reducing":[149],"asymmetry":[152],"gives":[154],"rise":[155],"aligning":[160],"inner-outer":[161],"behavior.":[162],"On":[163],"BABILong,":[164],"FocuSFT":[165,207],"improves":[166],"accuracy":[167],"by":[168,212],"up":[169],"+14pp":[171],"4K--32K":[173],"lengths;":[175],"RULER,":[177],"it":[178,195],"raises":[179],"CWE":[180],"aggregation":[181],"from":[182],"72.9\\%":[183],"81.1\\%":[185],"16K;":[187],"GPQA":[190],"with":[191],"agentic":[192],"tool":[193],"use,":[194],"yields":[196],"24\\%":[198],"relative":[199],"gain":[200],"pass@1.":[202],"Attention":[203],"analysis":[204],"shows":[205],"reduces":[208],"sink":[210],"mass":[211],"529$\\times$":[213],"triples":[215],"engagement":[217],"training.":[219],"Code:":[220],"https://github.com/JarvisPei/FocuSFT":[221]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-05-13T00:00:00"}