{"id":"https://openalex.org/W7224130108","doi":"https://doi.org/10.48550/arxiv.2610.37085","title":"On the Sparsity of Direct Preference Optimization: Weight Disentanglement in the NTK Regime","display_name":"On the Sparsity of Direct Preference Optimization: Weight Disentanglement in the NTK Regime","publication_year":2026,"publication_date":"2026-10-27","ids":{"openalex":"https://openalex.org/W7224130108","doi":"https://doi.org/10.48550/arxiv.2610.37085"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2610.37085","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.37085","pdf_url":"https://arxiv.org/pdf/2610.37085","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","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2610.37085","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5040755888","display_name":"Kensuke Sasaki","orcid":"https://orcid.org/0000-0002-6343-3669","observed_orcids":["https://orcid.org/0000-0002-6343-3669"]},"institutions":[{"id":"https://openalex.org/I74801974","display_name":"The University of Tokyo","ror":"https://ror.org/057zh3y96","country_code":"JP","type":"education","lineage":["https://openalex.org/I74801974"]}],"countries":["JP"],"is_corresponding":false,"raw_author_name":"Sasaki, Kensuke","raw_affiliation_strings":["Department of Computer Science The University of Tokyo"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Computer Science The University of Tokyo","institution_ids":["https://openalex.org/I74801974"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5060421432","display_name":"Issei Sato","orcid":"https://orcid.org/0000-0002-5066-1435","observed_orcids":["https://orcid.org/0000-0002-5066-1435"]},"institutions":[{"id":"https://openalex.org/I74801974","display_name":"The University of Tokyo","ror":"https://ror.org/057zh3y96","country_code":"JP","type":"education","lineage":["https://openalex.org/I74801974"]}],"countries":["JP"],"is_corresponding":false,"raw_author_name":"Sato, Issei","raw_affiliation_strings":["Department of Computer Science The University of Tokyo"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Computer Science The University of Tokyo","institution_ids":["https://openalex.org/I74801974"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I74801974"],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.35249999165534973,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.35249999165534973,"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":0.20020000636577606,"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/T11612","display_name":"Stochastic Gradient Optimization Techniques","score":0.10199999809265137,"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/direct-preference-optimization","display_name":"direct preference optimization","score":0.8489999771118164},{"id":"https://openalex.org/keywords/neural-tangent-kernel","display_name":"neural tangent kernel","score":0.7300000190734863},{"id":"https://openalex.org/keywords/support-vectors","display_name":"support vectors","score":0.6549999713897705},{"id":"https://openalex.org/keywords/model-merging","display_name":"model merging","score":0.8299999833106995},{"id":"https://openalex.org/keywords/reinforcement-learning-from-human-feedback","display_name":"reinforcement learning from human feedback","score":0.7900000214576721},{"id":"https://openalex.org/keywords/language-model-alignment","display_name":"language model alignment","score":0.6629999876022339}],"concepts":[{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.7156000137329102},{"id":"https://openalex.org/C74193536","wikidata":"https://www.wikidata.org/wiki/Q574844","display_name":"Kernel (algebra)","level":2,"score":0.5509999990463257},{"id":"https://openalex.org/C11210021","wikidata":"https://www.wikidata.org/wiki/Q1520713","display_name":"Linearization","level":3,"score":0.5006999969482422},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.47119998931884766},{"id":"https://openalex.org/C2781249084","wikidata":"https://www.wikidata.org/wiki/Q908656","display_name":"Preference","level":2,"score":0.47029998898506165},{"id":"https://openalex.org/C2780980858","wikidata":"https://www.wikidata.org/wiki/Q110022","display_name":"Dual (grammatical number)","level":2,"score":0.4293999969959259},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.4058000147342682},{"id":"https://openalex.org/C153258448","wikidata":"https://www.wikidata.org/wiki/Q1199743","display_name":"Gradient descent","level":3,"score":0.39399999380111694},{"id":"https://openalex.org/C2778029271","wikidata":"https://www.wikidata.org/wiki/Q5421931","display_name":"Extension (predicate logic)","level":2,"score":0.38269999623298645},{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.3702999949455261},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.3677999973297119},{"id":"https://openalex.org/C28826006","wikidata":"https://www.wikidata.org/wiki/Q33521","display_name":"Applied mathematics","level":1,"score":0.35260000824928284},{"id":"https://openalex.org/C124066611","wikidata":"https://www.wikidata.org/wiki/Q28684319","display_name":"Sparse approximation","level":2,"score":0.3138999938964844},{"id":"https://openalex.org/C2775997480","wikidata":"https://www.wikidata.org/wiki/Q586277","display_name":"Degree (music)","level":2,"score":0.3100000023841858},{"id":"https://openalex.org/C137836250","wikidata":"https://www.wikidata.org/wiki/Q984063","display_name":"Optimization problem","level":2,"score":0.3012000024318695},{"id":"https://openalex.org/C77553402","wikidata":"https://www.wikidata.org/wiki/Q13222579","display_name":"Upper and lower bounds","level":2,"score":0.295199990272522},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.2946999967098236},{"id":"https://openalex.org/C138187205","wikidata":"https://www.wikidata.org/wiki/Q131251","display_name":"Tangent","level":2,"score":0.2816999852657318},{"id":"https://openalex.org/C80884492","wikidata":"https://www.wikidata.org/wiki/Q3345678","display_name":"Reproducing kernel Hilbert space","level":3,"score":0.28130000829696655},{"id":"https://openalex.org/C12426560","wikidata":"https://www.wikidata.org/wiki/Q189569","display_name":"Basis (linear algebra)","level":2,"score":0.27880001068115234},{"id":"https://openalex.org/C83665646","wikidata":"https://www.wikidata.org/wiki/Q42139305","display_name":"Feature vector","level":2,"score":0.2766999900341034},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.2671999931335449},{"id":"https://openalex.org/C161584116","wikidata":"https://www.wikidata.org/wiki/Q1952580","display_name":"Multivariate statistics","level":2,"score":0.2624000012874603},{"id":"https://openalex.org/C147764199","wikidata":"https://www.wikidata.org/wiki/Q6865248","display_name":"Minification","level":2,"score":0.25679999589920044},{"id":"https://openalex.org/C18762648","wikidata":"https://www.wikidata.org/wiki/Q42213","display_name":"Work (physics)","level":2,"score":0.25290000438690186},{"id":"https://openalex.org/C149728462","wikidata":"https://www.wikidata.org/wiki/Q751319","display_name":"Minimax","level":2,"score":0.25220000743865967}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2610.37085","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.37085","pdf_url":"https://arxiv.org/pdf/2610.37085","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.2610.37085","doi":"https://doi.org/10.48550/arxiv.2610.37085","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2610.37085","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":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2610.37085","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.37085","pdf_url":"https://arxiv.org/pdf/2610.37085","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":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7224130108.pdf","grobid_xml":"https://content.openalex.org/works/W7224130108.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7204541737","https://openalex.org/W7139933476","https://openalex.org/W4414976198","https://openalex.org/W7104372316","https://openalex.org/W7160321001","https://openalex.org/W7220894715","https://openalex.org/W4414820685","https://openalex.org/W4402670121","https://openalex.org/W7123824242","https://openalex.org/W4403785627"],"abstract_inverted_index":{"Direct":[0],"Preference":[1],"Optimization":[2],"(DPO)":[3],"has":[4,18,34],"been":[5,35],"widely":[6],"adopted":[7],"for":[8,31,138],"aligning":[9],"language":[10],"models":[11,151],"with":[12,155],"human":[13],"preferences.":[14],"Recent":[15],"empirical":[16],"work":[17],"revealed":[19],"that":[20,50,172],"DPO":[21,56,77,173],"induces":[22],"remarkably":[23],"sparse":[24,82,92],"parameter":[25],"updates,":[26],"yet":[27],"no":[28],"theoretical":[29],"explanation":[30],"this":[32,67,109],"phenomenon":[33],"offered.":[36],"We":[37,107],"analyze":[38],"DPO's":[39],"optimization":[40],"dynamics":[41],"under":[42],"a":[43,81,122,128],"neural":[44],"tangent":[45],"kernel":[46],"linearization":[47],"and":[48,104,126,145,147,159,167],"show":[49,171],"the":[51,55,71,76,116,156],"logistic":[52],"structure":[53,110],"of":[54,73,91,118,165],"loss":[57],"gives":[58],"rise":[59],"to":[60,133,179,182],"an":[61],"implicit":[62],"maximum-margin":[63],"bias.":[64],"Building":[65],"on":[66,142,148,169],"observation,":[68],"we":[69],"introduce":[70],"concept":[72],"dual":[74,131],"sparsity:":[75],"update":[78],"direction":[79],"is":[80],"linear":[83],"combination":[84],"(data-level":[85],"sparsity,":[86,97,166],"driven":[87],"by":[88,121,177],"support":[89,119],"vectors)":[90],"gradient":[93,100],"feature":[94],"vectors":[95],"(parameter-level":[96],"arising":[98],"from":[99],"cancellation":[101],"between":[102],"preferred":[103],"dispreferred":[105],"responses).":[106],"formalize":[108],"in":[111],"our":[112],"main":[113],"theorem,":[114],"bound":[115],"number":[117],"pairs":[120],"margin-dependent":[123],"spectral":[124],"quantity,":[125],"present":[127],"framework":[129],"relating":[130],"sparsity":[132],"weight":[134],"disentanglement---a":[135],"condition":[136],"favorable":[137],"model":[139],"merging.":[140],"Experiments":[141],"GPT-2":[143,170],"Small":[144],"Medium":[146],"7--9B":[149],"instruction-tuned":[150],"(Llama-3-8B,":[152],"Mistral-7B,":[153],"Gemma-2-9B)":[154],"Anthropic":[157],"HH-RLHF":[158],"UltraFeedback":[160],"datasets":[161],"confirm":[162],"both":[163],"sources":[164],"experiments":[168],"reduces":[174],"cross-task":[175],"interference":[176],"up":[178],"$35{\\times}$":[180],"compared":[181],"supervised":[183],"fine-tuning.":[184]},"counts_by_year":[],"updated_date":"2026-10-31T12:09:11.096373","created_date":"2026-10-29T00:00:00"}