{"id":"https://openalex.org/W7131112085","doi":"https://doi.org/10.48550/arxiv.2602.17835","title":"Influence-Preserving Proxies for Gradient-Based Data Selection in LLM Fine-tuning","display_name":"Influence-Preserving Proxies for Gradient-Based Data Selection in LLM Fine-tuning","publication_year":2026,"publication_date":"2026-02-19","ids":{"openalex":"https://openalex.org/W7131112085","doi":"https://doi.org/10.48550/arxiv.2602.17835"},"language":null,"primary_location":{"id":"pmh:doi:10.48550/arxiv.2602.17835","is_oa":true,"landing_page_url":null,"pdf_url":null,"source":{"id":"https://openalex.org/S4406922384","display_name":"Open MIND","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Article"},"type":"article","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":null,"any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5126654753","display_name":"Sirui Chen","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chen, Sirui","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5126592657","display_name":"Yunzhe Qi","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Qi, Yunzhe","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5110060470","display_name":"M. Ai","orcid":"https://orcid.org/0009-0009-3878-6188"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ai, Mengting","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5126643475","display_name":"Yifan Sun","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Sun, Yifan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5126640655","display_name":"Ruizhong Qiu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Qiu, Ruizhong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5051330767","display_name":"Jiaru Zou","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zou, Jiaru","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5126600469","display_name":"Jingrui He","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"He, Jingrui","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":0.0,"has_fulltext":false,"cited_by_count":0,"citation_normalized_percentile":{"value":0.20657265,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"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":0.40849998593330383,"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":0.40849998593330383,"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":0.19349999725818634,"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/T10201","display_name":"Speech Recognition and Synthesis","score":0.07660000026226044,"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/leverage","display_name":"Leverage (statistics)","score":0.745199978351593},{"id":"https://openalex.org/keywords/selection","display_name":"Selection (genetic algorithm)","score":0.5216000080108643},{"id":"https://openalex.org/keywords/proxy","display_name":"Proxy (statistics)","score":0.5102999806404114},{"id":"https://openalex.org/keywords/model-selection","display_name":"Model selection","score":0.46230000257492065},{"id":"https://openalex.org/keywords/scalability","display_name":"Scalability","score":0.43799999356269836},{"id":"https://openalex.org/keywords/feature-selection","display_name":"Feature selection","score":0.3912000060081482}],"concepts":[{"id":"https://openalex.org/C153083717","wikidata":"https://www.wikidata.org/wiki/Q6535263","display_name":"Leverage (statistics)","level":2,"score":0.745199978351593},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7235000133514404},{"id":"https://openalex.org/C81917197","wikidata":"https://www.wikidata.org/wiki/Q628760","display_name":"Selection (genetic algorithm)","level":2,"score":0.5216000080108643},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.5109000205993652},{"id":"https://openalex.org/C2780148112","wikidata":"https://www.wikidata.org/wiki/Q1432581","display_name":"Proxy (statistics)","level":2,"score":0.5102999806404114},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.48489999771118164},{"id":"https://openalex.org/C93959086","wikidata":"https://www.wikidata.org/wiki/Q6888345","display_name":"Model selection","level":2,"score":0.46230000257492065},{"id":"https://openalex.org/C48044578","wikidata":"https://www.wikidata.org/wiki/Q727490","display_name":"Scalability","level":2,"score":0.43799999356269836},{"id":"https://openalex.org/C148483581","wikidata":"https://www.wikidata.org/wiki/Q446488","display_name":"Feature selection","level":2,"score":0.3912000060081482},{"id":"https://openalex.org/C12725497","wikidata":"https://www.wikidata.org/wiki/Q810247","display_name":"Baseline (sea)","level":2,"score":0.3790000081062317},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3522000014781952},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.2827000021934509},{"id":"https://openalex.org/C2779714256","wikidata":"https://www.wikidata.org/wiki/Q25305062","display_name":"Multiple Models","level":2,"score":0.28060001134872437},{"id":"https://openalex.org/C160920958","wikidata":"https://www.wikidata.org/wiki/Q7662746","display_name":"Synthetic data","level":2,"score":0.26030001044273376},{"id":"https://openalex.org/C45804977","wikidata":"https://www.wikidata.org/wiki/Q7239673","display_name":"Predictive modelling","level":2,"score":0.2565000057220459}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:doi:10.48550/arxiv.2602.17835","is_oa":true,"landing_page_url":null,"pdf_url":null,"source":{"id":"https://openalex.org/S4406922384","display_name":"Open MIND","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Article"},{"id":"doi:10.48550/arxiv.2602.17835","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2602.17835","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":"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:doi:10.48550/arxiv.2602.17835","is_oa":true,"landing_page_url":null,"pdf_url":null,"source":{"id":"https://openalex.org/S4406922384","display_name":"Open MIND","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Article"},"sustainable_development_goals":[{"score":0.4165406823158264,"display_name":"Quality Education","id":"https://metadata.un.org/sdg/4"}],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Supervised":[0],"fine-tuning":[1],"(SFT)":[2],"relies":[3],"critically":[4],"on":[5,187],"selecting":[6],"training":[7],"data":[8,17,220],"that":[9,100,139,160,212],"most":[10],"benefits":[11],"a":[12,97,112,171],"model's":[13,148],"downstream":[14],"performance.":[15],"Gradient-based":[16],"selection":[18,221],"methods":[19],"such":[20],"as":[21,56],"TracIn":[22],"and":[23,74,124,134,156,166],"Influence":[24],"Functions":[25],"leverage":[26],"influence":[27,88,118],"to":[28,51,116,129,204],"identify":[29],"useful":[30],"samples,":[31],"but":[32,58],"their":[33,63,68],"computational":[34,142,197],"cost":[35,143,198],"scales":[36],"poorly,":[37],"making":[38,218],"them":[39],"impractical":[40],"for":[41,224],"multi-billion-parameter":[42],"large":[43],"language":[44],"models":[45,55],"(LLMs).":[46],"A":[47],"common":[48],"alternative":[49],"is":[50],"use":[52],"off-the-shelf":[53,164,184],"smaller":[54],"proxies,":[57,217],"they":[59,75],"remain":[60],"suboptimal":[61],"since":[62],"learning":[64],"dynamics":[65],"are":[66],"unclear,":[67],"sizes":[69],"cannot":[70,76],"be":[71,77],"flexibly":[72,140],"adjusted,":[73],"further":[78],"aligned":[79],"with":[80,175],"the":[81,106,121,146,181,205],"target":[82,107,122,147],"model":[83,132],"in":[84],"terms":[85],"of":[86,120],"gradient-based":[87,219],"estimation.":[89],"To":[90],"address":[91],"these":[92],"challenges,":[93],"we":[94],"introduce":[95],"Iprox,":[96],"two-stage":[98],"framework":[99],"derives":[101],"influence-preserving":[102,216],"proxies":[103,138,165],"directly":[104],"from":[105],"model.":[108,208],"It":[109],"first":[110],"applies":[111],"low-rank":[113],"compression":[114],"stage":[115,128],"preserve":[117],"information":[119],"model,":[123],"then":[125],"an":[126],"aligning":[127],"align":[130],"both":[131],"gradients":[133],"logits,":[135],"thereby":[136],"constructing":[137],"control":[141],"while":[144,195],"retaining":[145],"influence.":[149],"Experimental":[150],"results":[151,210],"across":[152],"diverse":[153],"LLM":[154],"families":[155],"evaluation":[157],"tasks":[158],"show":[159,211],"Iprox":[161,176,189,213],"consistently":[162],"outperforms":[163],"baseline":[167],"methods.":[168],"On":[169],"Qwen3-4B,":[170],"1.5B":[172],"proxy":[173],"constructed":[174],"achieves":[177,190],"stronger":[178],"performance":[179,192],"than":[180,193,201],"larger":[182],"1.7B":[183],"proxy.":[185],"Notably,":[186],"Llama3.2,":[188],"better":[191],"baselines":[194],"reducing":[196],"by":[199],"more":[200,222],"half":[202],"relative":[203],"full":[206],"3B":[207],"These":[209],"provides":[214],"effective":[215],"scalable":[223],"LLMs.":[225]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-02-24T00:00:00"}