{"id":"https://openalex.org/W7223337814","doi":"https://doi.org/10.48550/arxiv.2610.30934","title":"Verbalizing Activation Differences with Natural Language Autoencoders","display_name":"Verbalizing Activation Differences with Natural Language Autoencoders","publication_year":2026,"publication_date":"2026-10-22","ids":{"openalex":"https://openalex.org/W7223337814","doi":"https://doi.org/10.48550/arxiv.2610.30934"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2610.30934","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.30934","pdf_url":"https://arxiv.org/pdf/2610.30934","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","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2610.30934","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5159365088","display_name":"Andrea Blasi N\u00fa\u00f1ez","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I177969490","display_name":"University of Southern Denmark","ror":"https://ror.org/03yrrjy16","country_code":"DK","type":"education","lineage":["https://openalex.org/I177969490"]}],"countries":["DK"],"is_corresponding":false,"raw_author_name":"N\u00fa\u00f1ez, Andrea Blasi","raw_affiliation_strings":["University of Southern Denmark"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Southern Denmark","institution_ids":["https://openalex.org/I177969490"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5120421625","display_name":"Gianluca Barmina","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I177969490","display_name":"University of Southern Denmark","ror":"https://ror.org/03yrrjy16","country_code":"DK","type":"education","lineage":["https://openalex.org/I177969490"]}],"countries":["DK"],"is_corresponding":false,"raw_author_name":"Barmina, Gianluca","raw_affiliation_strings":["University of Southern Denmark"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Southern Denmark","institution_ids":["https://openalex.org/I177969490"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5023712450","display_name":"Federico Torrielli","orcid":"https://orcid.org/0000-0001-8037-8828","observed_orcids":["https://orcid.org/0000-0001-8037-8828"]},"institutions":[{"id":"https://openalex.org/I177969490","display_name":"University of Southern Denmark","ror":"https://ror.org/03yrrjy16","country_code":"DK","type":"education","lineage":["https://openalex.org/I177969490"]}],"countries":["DK"],"is_corresponding":false,"raw_author_name":"Torrielli, Federico","raw_affiliation_strings":["University of Southern Denmark"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Southern Denmark","institution_ids":["https://openalex.org/I177969490"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5069935355","display_name":"Thiusius Rajeeth Savarimuthu","orcid":"https://orcid.org/0000-0002-2478-8694","observed_orcids":["https://orcid.org/0000-0002-2478-8694"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Savarimuthu, Thiusius Rajeeth","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5058200123","display_name":"Peter Schneider\u2013Kamp","orcid":"https://orcid.org/0000-0003-4000-5570","observed_orcids":["https://orcid.org/0000-0003-4000-5570"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Schneider-Kamp, Peter","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5138298884","display_name":"Lukas Galke Poech","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Poech, Lukas Galke","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.8300999999046326,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.8300999999046326,"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":0.1289999932050705,"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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.02449999935925007,"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/large-language-models","display_name":"large language models","score":0.722000002861023},{"id":"https://openalex.org/keywords/misalignment-detection","display_name":"misalignment detection","score":0.6869999766349792}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.830299973487854},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.6183000206947327},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.5806000232696533},{"id":"https://openalex.org/C195324797","wikidata":"https://www.wikidata.org/wiki/Q33742","display_name":"Natural language","level":2,"score":0.5748999714851379},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5726000070571899},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.5264000296592712},{"id":"https://openalex.org/C2779439875","wikidata":"https://www.wikidata.org/wiki/Q1078276","display_name":"Natural language understanding","level":3,"score":0.4812999963760376},{"id":"https://openalex.org/C205711294","wikidata":"https://www.wikidata.org/wiki/Q176953","display_name":"Rendering (computer graphics)","level":2,"score":0.4016999900341034},{"id":"https://openalex.org/C521332185","wikidata":"https://www.wikidata.org/wiki/Q185816","display_name":"Analogy","level":2,"score":0.3944999873638153},{"id":"https://openalex.org/C2776608160","wikidata":"https://www.wikidata.org/wiki/Q4785462","display_name":"Natural (archaeology)","level":2,"score":0.365200012922287},{"id":"https://openalex.org/C2776187449","wikidata":"https://www.wikidata.org/wiki/Q1513879","display_name":"Natural language generation","level":3,"score":0.35679998993873596},{"id":"https://openalex.org/C2983448237","wikidata":"https://www.wikidata.org/wiki/Q1078276","display_name":"Language understanding","level":2,"score":0.3547999858856201},{"id":"https://openalex.org/C94922259","wikidata":"https://www.wikidata.org/wiki/Q33215","display_name":"Constructed language","level":2,"score":0.3521000146865845},{"id":"https://openalex.org/C2776323365","wikidata":"https://www.wikidata.org/wiki/Q171180","display_name":"Taboo","level":2,"score":0.3513000011444092},{"id":"https://openalex.org/C2777601683","wikidata":"https://www.wikidata.org/wiki/Q6499736","display_name":"Vocabulary","level":2,"score":0.3481999933719635},{"id":"https://openalex.org/C107457646","wikidata":"https://www.wikidata.org/wiki/Q207434","display_name":"Human\u2013computer interaction","level":1,"score":0.3409000039100647},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.31310001015663147},{"id":"https://openalex.org/C184337299","wikidata":"https://www.wikidata.org/wiki/Q1437428","display_name":"Semantics (computer science)","level":2,"score":0.3082999885082245},{"id":"https://openalex.org/C2779530757","wikidata":"https://www.wikidata.org/wiki/Q1207505","display_name":"Quality (philosophy)","level":2,"score":0.302700012922287},{"id":"https://openalex.org/C2780922921","wikidata":"https://www.wikidata.org/wiki/Q255189","display_name":"Paraphrase","level":2,"score":0.27399998903274536},{"id":"https://openalex.org/C90805587","wikidata":"https://www.wikidata.org/wiki/Q10944557","display_name":"Word (group theory)","level":2,"score":0.2630000114440918},{"id":"https://openalex.org/C180747234","wikidata":"https://www.wikidata.org/wiki/Q23373","display_name":"Cognitive psychology","level":1,"score":0.2524000108242035}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2610.30934","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.30934","pdf_url":"https://arxiv.org/pdf/2610.30934","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.2610.30934","doi":"https://doi.org/10.48550/arxiv.2610.30934","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2610.30934","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2610.30934","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.30934","pdf_url":"https://arxiv.org/pdf/2610.30934","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":[{"id":"https://openalex.org/F4320314239","display_name":"Netherlands eScience Center","ror":"https://ror.org/00rbjv475"},{"id":"https://openalex.org/F4320322436","display_name":"Novo Nordisk","ror":"https://ror.org/0435rc536"},{"id":"https://openalex.org/F4320323799","display_name":"Syddansk Universitet","ror":"https://ror.org/03yrrjy16"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7223337814.pdf","grobid_xml":"https://content.openalex.org/works/W7223337814.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7221825461","https://openalex.org/W7214788809","https://openalex.org/W7162394530","https://openalex.org/W7214805248","https://openalex.org/W7170190251","https://openalex.org/W4415317977","https://openalex.org/W6910022981","https://openalex.org/W7116043699","https://openalex.org/W7204640880","https://openalex.org/W7202179404"],"abstract_inverted_index":{"Activation":[0],"verbalizers":[1],"translate":[2],"the":[3,40,78,128,133,158,173,200,204,211,217,230],"internal":[4],"representations":[5],"of":[6,77,122,145],"large":[7],"language":[8,17,28],"models":[9,103,171],"into":[10],"human-readable":[11],"text.":[12],"The":[13],"recently":[14],"proposed":[15],"natural":[16,27],"autoencoders":[18],"(NLAs)":[19],"learn":[20],"this":[21,49,188],"by":[22],"reconstructing":[23],"activations":[24,125],"through":[25],"a":[26,112,237,254],"bottleneck.":[29],"However,":[30],"it":[31],"remains":[32],"unclear":[33],"how":[34],"well":[35],"these":[36],"methods":[37,228],"generalize":[38,73],"beyond":[39],"setting":[41],"they":[42,80],"were":[43,81],"trained":[44,82],"for.":[45],"Here,":[46],"we":[47],"study":[48],"question":[50],"for":[51,98,157,177,257],"NLAs":[52,72,86,95,197,220,244,253],"along":[53],"four":[54,102,118],"closely":[55],"related":[56],"questions:":[57],"(i)":[58],"Do":[59,71],"NLA":[60,138],"verbalizations":[61,121],"exhibit":[62],"compositional":[63],"structure":[64],"with":[65,151],"respect":[66],"to":[67,74,116],"their":[68,169],"input?":[69],"(ii)":[70],"finetuned":[75,205],"versions":[76],"model":[79,91,99,148,166,206,226],"for?":[83],"(iii)":[84],"Can":[85,94],"recover":[87,199],"finetuning":[88,201],"objectives":[89],"in":[90,249],"diffing?":[92,100],"(iv)":[93],"be":[96,140],"specialized":[97],"Across":[101],"from":[104],"three":[105],"families,":[106],"our":[107],"experiments":[108],"provide":[109],"evidence":[110],"supporting":[111],"largely":[113],"positive":[114],"answer":[115],"all":[117],"questions.":[119],"Specifically,":[120],"composed":[123],"analogy":[124],"mostly":[126],"name":[127],"target":[129],"while":[130],"rarely":[131],"mentioning":[132],"subtracted":[134],"word,":[135],"and":[136,168,180],"an":[137],"can":[139,198,245],"reused":[141],"on":[142,183,190,229],"lightweight":[143],"finetunes":[144],"its":[146],"original":[147,170],"without":[149,203],"retraining,":[150],"reconstruction":[152],"quality":[153],"degrading":[154],"noticeably":[155],"only":[156],"largest":[159],"model.":[160],"Verbalizing":[161],"activation":[162,184,250],"differences":[163,185,248],"between":[164],"Taboo":[165],"organisms":[167],"recovers":[172],"concealed":[174],"secret":[175],"word":[176],"most":[178,212],"organisms,":[179],"reinforcement":[181],"learning":[182],"further":[186],"improves":[187],"recovery":[189],"held-out":[191],"organisms.":[192],"We":[193],"also":[194],"find":[195],"that":[196,243],"objective":[202],"generating":[207],"any":[208],"token,":[209],"since":[210],"informative":[213],"token":[214],"positions":[215],"precede":[216],"user":[218],"prompt.":[219],"do":[221],"not":[222],"yet":[223],"match":[224],"state-of-the-art":[225],"diffing":[227],"emergent":[231],"misalignment":[232],"detection":[233],"task,":[234],"but":[235],"outperform":[236],"blackbox":[238],"baseline.":[239],"Our":[240],"findings":[241],"suggest":[242],"meaningfully":[246],"verbalize":[247],"space,":[251],"rendering":[252],"promising":[255],"direction":[256],"future":[258],"work.":[259]},"counts_by_year":[],"updated_date":"2026-10-28T10:33:24.094632","created_date":"2026-10-24T00:00:00"}