{"id":"https://openalex.org/W4396822045","doi":"https://doi.org/10.48550/arxiv.2405.05466","title":"Poser: Unmasking Alignment Faking LLMs by Manipulating Their Internals","display_name":"Poser: Unmasking Alignment Faking LLMs by Manipulating Their Internals","publication_year":2024,"publication_date":"2024-05-08","ids":{"openalex":"https://openalex.org/W4396822045","doi":"https://doi.org/10.48550/arxiv.2405.05466"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2405.05466","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2405.05466","pdf_url":"https://arxiv.org/pdf/2405.05466","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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":"","raw_type":"text"},"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2405.05466","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5097450395","display_name":"Joshua Clymer","orcid":null},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Clymer, Joshua","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5097450396","display_name":"Caden Juang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Juang, Caden","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5097424532","display_name":"Severin Field","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Field, Severin","raw_affiliation_strings":[],"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":3,"corresponding_author_ids":["https://openalex.org/A5097450395"],"corresponding_institution_ids":[],"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/T11652","display_name":"Imbalanced Data Classification Techniques","score":0.9681000113487244,"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/T11652","display_name":"Imbalanced Data Classification Techniques","score":0.9681000113487244,"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/business","display_name":"Business","score":0.4021882712841034},{"id":"https://openalex.org/keywords/computer-science","display_name":"Computer science","score":0.393218457698822}],"concepts":[{"id":"https://openalex.org/C144133560","wikidata":"https://www.wikidata.org/wiki/Q4830453","display_name":"Business","level":0,"score":0.4021882712841034},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.393218457698822}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2405.05466","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2405.05466","pdf_url":"https://arxiv.org/pdf/2405.05466","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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":"","raw_type":"text"},{"id":"doi:10.48550/arxiv.2405.05466","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2405.05466","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":"article"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2405.05466","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2405.05466","pdf_url":"https://arxiv.org/pdf/2405.05466","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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":"","raw_type":"text"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4396822045.pdf","grobid_xml":"https://content.openalex.org/works/W4396822045.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4391375266","https://openalex.org/W2748952813","https://openalex.org/W2390279801","https://openalex.org/W2358668433","https://openalex.org/W4396701345","https://openalex.org/W2376932109","https://openalex.org/W2001405890","https://openalex.org/W4396696052","https://openalex.org/W2382290278","https://openalex.org/W4395014643"],"abstract_inverted_index":{"Like":[0],"a":[1,21,38],"criminal":[2],"under":[3],"investigation,":[4],"Large":[5],"Language":[6],"Models":[7],"(LLMs)":[8],"might":[9],"pretend":[10],"to":[11,48,73,81],"be":[12,74],"aligned":[13],"while":[14],"evaluated":[15],"and":[16],"misbehave":[17],"when":[18],"they":[19],"have":[20],"good":[22],"opportunity.":[23],"Can":[24],"current":[25],"interpretability":[26],"methods":[27],"catch":[28],"these":[29],"'alignment":[30],"fakers?'":[31],"To":[32],"answer":[33],"this":[34],"question,":[35],"we":[36],"introduce":[37],"benchmark":[39],"that":[40],"consists":[41],"of":[42,45,102,106],"324":[43],"pairs":[44],"LLMs":[46],"fine-tuned":[47],"select":[49],"actions":[50],"in":[51,56,67],"role-play":[52],"scenarios.":[53],"One":[54],"model":[55,65,86],"each":[57],"pair":[58],"is":[59,71,80],"consistently":[60],"benign":[61],"(aligned).":[62],"The":[63,78],"other":[64],"misbehaves":[66],"scenarios":[68],"where":[69,90],"it":[70],"unlikely":[72],"caught":[75],"(alignment":[76],"faking).":[77],"task":[79],"identify":[82],"the":[83,91],"alignment":[84],"faking":[85],"using":[87],"only":[88],"inputs":[89],"two":[92],"models":[93],"behave":[94],"identically.":[95],"We":[96],"test":[97],"five":[98],"detection":[99],"strategies,":[100],"one":[101],"which":[103],"identifies":[104],"98%":[105],"alignment-fakers.":[107]},"counts_by_year":[],"updated_date":"2026-03-09T08:58:05.943551","created_date":"2025-10-10T00:00:00"}