{"id":"https://openalex.org/W7163912613","doi":"https://doi.org/10.48550/arxiv.2606.06667","title":"The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignment","display_name":"The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignment","publication_year":2026,"publication_date":"2026-06-04","ids":{"openalex":"https://openalex.org/W7163912613","doi":"https://doi.org/10.48550/arxiv.2606.06667"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.06667","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.06667","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":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.2606.06667","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5029412635","display_name":"Jiachen Zhao","orcid":"https://orcid.org/0000-0002-2864-9229"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhao, Jiachen","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5102822963","display_name":"Zhengxuan Wu","orcid":"https://orcid.org/0000-0001-5581-8908"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wu, Zhengxuan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5082261951","display_name":"Aryaman Arora","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Arora, Aryaman","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138125255","display_name":"Yiyou Sun","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Sun, Yiyou","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138194158","display_name":"David Bau","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Bau, David","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5138158117","display_name":"Weiyan Shi","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Shi, Weiyan","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":0.5727999806404114,"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":0.5727999806404114,"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":0.23000000417232513,"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/T13629","display_name":"Text Readability and Simplification","score":0.05260000005364418,"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/set","display_name":"Set (abstract data type)","score":0.6313999891281128},{"id":"https://openalex.org/keywords/interleaving","display_name":"Interleaving","score":0.6062999963760376},{"id":"https://openalex.org/keywords/prefix","display_name":"Prefix","score":0.553600013256073},{"id":"https://openalex.org/keywords/generalization","display_name":"Generalization","score":0.5529999732971191},{"id":"https://openalex.org/keywords/path","display_name":"Path (computing)","score":0.43650001287460327},{"id":"https://openalex.org/keywords/unintended-consequences","display_name":"Unintended consequences","score":0.4088999927043915},{"id":"https://openalex.org/keywords/boosting","display_name":"Boosting (machine learning)","score":0.40049999952316284},{"id":"https://openalex.org/keywords/security-token","display_name":"Security token","score":0.39800000190734863}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7897999882698059},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.6313999891281128},{"id":"https://openalex.org/C28034677","wikidata":"https://www.wikidata.org/wiki/Q17092530","display_name":"Interleaving","level":2,"score":0.6062999963760376},{"id":"https://openalex.org/C141603448","wikidata":"https://www.wikidata.org/wiki/Q134830","display_name":"Prefix","level":2,"score":0.553600013256073},{"id":"https://openalex.org/C177148314","wikidata":"https://www.wikidata.org/wiki/Q170084","display_name":"Generalization","level":2,"score":0.5529999732971191},{"id":"https://openalex.org/C2777735758","wikidata":"https://www.wikidata.org/wiki/Q817765","display_name":"Path (computing)","level":2,"score":0.43650001287460327},{"id":"https://openalex.org/C2776889888","wikidata":"https://www.wikidata.org/wiki/Q1135789","display_name":"Unintended consequences","level":2,"score":0.4088999927043915},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.40560001134872437},{"id":"https://openalex.org/C46686674","wikidata":"https://www.wikidata.org/wiki/Q466303","display_name":"Boosting (machine learning)","level":2,"score":0.40049999952316284},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.39800000190734863},{"id":"https://openalex.org/C201943243","wikidata":"https://www.wikidata.org/wiki/Q7551008","display_name":"Social connectedness","level":2,"score":0.3905999958515167},{"id":"https://openalex.org/C107457646","wikidata":"https://www.wikidata.org/wiki/Q207434","display_name":"Human\u2013computer interaction","level":1,"score":0.3693000078201294},{"id":"https://openalex.org/C18762648","wikidata":"https://www.wikidata.org/wiki/Q42213","display_name":"Work (physics)","level":2,"score":0.3601999878883362},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.351500004529953},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.3440000116825104},{"id":"https://openalex.org/C38652104","wikidata":"https://www.wikidata.org/wiki/Q3510521","display_name":"Computer security","level":1,"score":0.335999995470047},{"id":"https://openalex.org/C100660578","wikidata":"https://www.wikidata.org/wiki/Q18733","display_name":"Recall","level":2,"score":0.3269999921321869},{"id":"https://openalex.org/C2777267654","wikidata":"https://www.wikidata.org/wiki/Q3519023","display_name":"Test (biology)","level":2,"score":0.29350000619888306},{"id":"https://openalex.org/C98184364","wikidata":"https://www.wikidata.org/wiki/Q1780131","display_name":"Argument (complex analysis)","level":2,"score":0.28679999709129333},{"id":"https://openalex.org/C193702766","wikidata":"https://www.wikidata.org/wiki/Q1414548","display_name":"Concurrency","level":2,"score":0.2766000032424927},{"id":"https://openalex.org/C165696696","wikidata":"https://www.wikidata.org/wiki/Q11287","display_name":"Exploit","level":2,"score":0.27399998903274536},{"id":"https://openalex.org/C111335779","wikidata":"https://www.wikidata.org/wiki/Q3454686","display_name":"Reduction (mathematics)","level":2,"score":0.27070000767707825},{"id":"https://openalex.org/C71559656","wikidata":"https://www.wikidata.org/wiki/Q671298","display_name":"Divide and conquer algorithms","level":2,"score":0.25459998846054077},{"id":"https://openalex.org/C142944206","wikidata":"https://www.wikidata.org/wiki/Q1786137","display_name":"Proactivity","level":2,"score":0.2538999915122986}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.06667","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.06667","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":"doi:10.48550/arxiv.2606.06667","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.06667","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[{"id":"https://metadata.un.org/sdg/16","display_name":"Peace, Justice and strong institutions","score":0.4666886031627655}],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"The":[0],"mechanisms":[1],"behind":[2],"LLMs'":[3],"broad":[4,24],"over-generalization":[5],"beyond":[6],"training":[7,100],"examples":[8],"remain":[9],"unclear.":[10],"Emergent":[11],"misalignment":[12,25],"(EM)":[13],"offers":[14],"a":[15,135,184],"striking":[16],"case":[17],"study:":[18],"finetuning":[19],"on":[20,86,121,163],"narrow":[21],"tasks":[22],"induces":[23],"to":[26,58,101,147],"semantically-unrelated":[27],"test":[28],"domains.":[29,206],"In":[30],"this":[31,51,87],"work,":[32],"we":[33,89],"propose":[34,90],"the":[35,38,43,59,68,74,82,122,166],"Piggyback":[36,167],"Hypothesis:":[37],"chat-template":[39],"tokens":[40],"can":[41,77,201],"piggyback":[42,202],"finetuned":[44,120],"behaviour":[45],"onto":[46],"out-of-domain":[47],"queries.":[48],"We":[49,141],"validate":[50],"hypothesis":[52],"by":[53,161],"showing":[54],"that":[55,144,173],"subtle":[56],"perturbations":[57],"prefix":[60,69],"(tokens":[61],"preceding":[62],"all":[63],"user":[64,83],"queries),":[65],"or":[66],"patching":[67],"representations":[70,98],"with":[71,134],"those":[72],"from":[73],"unfinetuned":[75],"model,":[76],"restore":[78],"alignment":[79],"without":[80],"changing":[81],"query.":[84],"Building":[85],"finding,":[88],"Token-Regularized":[91],"Finetuning":[92],"(TReFT),":[93],"which":[94],"regularizes":[95],"specific":[96],"token":[97],"during":[99],"mitigate":[102],"EM.":[103],"Across":[104],"different":[105],"models":[106],"and":[107,155,177,182],"multiple":[108],"EM-inducing":[109],"datasets,":[110],"TReFT":[111,125,145],"reduces":[112],"EM":[113,129],"while":[114],"preserving":[115],"in-domain":[116],"learning.":[117],"On":[118],"Llama-3.1-8B":[119],"legal":[123],"domain,":[124],"achieves":[126],"33.5%":[127],"more":[128,187],"reduction":[130],"than":[131],"data":[132],"interleaving":[133],"retain":[136],"set":[137],"of":[138,196],"aligned":[139],"examples.":[140],"further":[142,194],"show":[143],"extends":[146],"other":[148],"narrow-finetuning":[149],"settings,":[150],"including":[151],"abstention,":[152],"tool":[153],"use,":[154],"refusal":[156],"(off-topic":[157],"generalization":[158],"is":[159],"reduced":[160],"54.3%":[162],"average),":[164],"supporting":[165],"Hypothesis.":[168],"Broadly,":[169],"our":[170],"work":[171],"highlights":[172],"LLMs":[174],"may":[175],"learn":[176],"generalize":[178],"in":[179],"unintended":[180],"ways":[181],"suggests":[183],"path":[185],"toward":[186],"constrained":[188],"finetuning.":[189],"It":[190],"also":[191],"calls":[192],"for":[193],"study":[195],"how":[197],"shared":[198],"input":[199],"features":[200],"model":[203],"behavior":[204],"across":[205]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-06-09T00:00:00"}