{"id":"https://openalex.org/W4403800322","doi":"https://doi.org/10.48550/arxiv.2409.17113","title":"Characterizing stable regions in the residual stream of LLMs","display_name":"Characterizing stable regions in the residual stream of LLMs","publication_year":2024,"publication_date":"2024-09-25","ids":{"openalex":"https://openalex.org/W4403800322","doi":"https://doi.org/10.48550/arxiv.2409.17113"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2409.17113","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2409.17113","pdf_url":"https://arxiv.org/pdf/2409.17113","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/2409.17113","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5114412486","display_name":"Jett Janiak","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Janiak, Jett","raw_affiliation_strings":["LASR Labs"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"LASR Labs","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5121764602","display_name":"Jacek Karwowski","orcid":"https://orcid.org/0000-0002-8361-2912","observed_orcids":["https://orcid.org/0000-0002-8361-2912"]},"institutions":[{"id":"https://openalex.org/I40120149","display_name":"University of Oxford","ror":"https://ror.org/052gg0110","country_code":"GB","type":"education","lineage":["https://openalex.org/I40120149"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Karwowski, Jacek","raw_affiliation_strings":["University of Oxford"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Oxford","institution_ids":["https://openalex.org/I40120149"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5114412485","display_name":"Chatrik Singh Mangat","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Mangat, Chatrik Singh","raw_affiliation_strings":["LASR Labs"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"LASR Labs","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5114412483","display_name":"Giorgi Giglemiani","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Giglemiani, Giorgi","raw_affiliation_strings":["LASR Labs"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"LASR Labs","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5114412484","display_name":"Nora Petrova","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Petrova, Nora","raw_affiliation_strings":["LASR Labs"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"LASR Labs","institution_ids":[]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5114412487","display_name":"Stefan Heimersheim","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Heimersheim, Stefan","raw_affiliation_strings":["Apollo Research"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Apollo Research","institution_ids":[]}]}],"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":6,"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":0.9707000255584717,"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":0.9707000255584717,"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":0.011599999852478504,"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":0.011500000022351742,"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/residual-stream","display_name":"residual stream","score":0.6990000009536743},{"id":"https://openalex.org/keywords/transformer","display_name":"transformer models","score":0.7099999785423279},{"id":"https://openalex.org/keywords/stable-regions","display_name":"stable regions","score":0.8069999814033508},{"id":"https://openalex.org/keywords/activation-space","display_name":"activation space","score":0.5559999942779541},{"id":"https://openalex.org/keywords/polytopes","display_name":"polytopes","score":0.6380000114440918},{"id":"https://openalex.org/keywords/training-dynamics","display_name":"training dynamics","score":0.6389999985694885},{"id":"https://openalex.org/keywords/model-size","display_name":"model size","score":0.45100000500679016},{"id":"https://openalex.org/keywords/next-token-prediction","display_name":"next-token prediction","score":0.9190000295639038},{"id":"https://openalex.org/keywords/interpretability","display_name":"interpretability","score":0.8429999947547913},{"id":"https://openalex.org/keywords/neural-network-complexity","display_name":"neural network complexity","score":0.8669999837875366}],"concepts":[{"id":"https://openalex.org/C155512373","wikidata":"https://www.wikidata.org/wiki/Q287450","display_name":"Residual","level":2,"score":0.6866298913955688},{"id":"https://openalex.org/C205649164","wikidata":"https://www.wikidata.org/wiki/Q1071","display_name":"Geography","level":0,"score":0.34632408618927},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.23257508873939514},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.09518563747406006}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2409.17113","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2409.17113","pdf_url":"https://arxiv.org/pdf/2409.17113","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.2409.17113","doi":"https://doi.org/10.48550/arxiv.2409.17113","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2409.17113","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:2409.17113","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2409.17113","pdf_url":"https://arxiv.org/pdf/2409.17113","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/F4320315389","display_name":"Open Philanthropy Project","ror":"https://ror.org/004d1k391"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4403800322.pdf","grobid_xml":"https://content.openalex.org/works/W4403800322.grobid-xml"},"referenced_works_count":5,"referenced_works":["https://openalex.org/W4296932880","https://openalex.org/W4309874180","https://openalex.org/W4402671236","https://openalex.org/W4402724519","https://openalex.org/W4403780485"],"related_works":["https://openalex.org/W7128497383","https://openalex.org/W4406231212","https://openalex.org/W7161278272","https://openalex.org/W4415797094","https://openalex.org/W7203498737","https://openalex.org/W7164038692","https://openalex.org/W7214893629","https://openalex.org/W7164846351","https://openalex.org/W4407759155","https://openalex.org/W7130675575"],"abstract_inverted_index":{"We":[0],"identify":[1],"stable":[2,59],"regions":[3,28,44,60],"in":[4],"the":[5,11,74,92],"residual":[6],"stream":[7],"of":[8,94],"Transformers,":[9],"where":[10,65],"model's":[12],"output":[13],"remains":[14],"insensitive":[15],"to":[16,46,78],"small":[17],"activation":[18],"changes,":[19],"but":[20],"exhibits":[21],"high":[22],"sensitivity":[23],"at":[24],"region":[25,76],"boundaries.":[26],"These":[27],"emerge":[29],"during":[30],"training":[31,37,100],"and":[32,71,102],"become":[33],"more":[34],"defined":[35],"as":[36],"progresses":[38],"or":[39],"model":[40],"size":[41],"increases.":[42],"The":[43],"appear":[45],"be":[47],"much":[48],"larger":[49],"than":[50],"previously":[51],"studied":[52],"polytopes.":[53],"Our":[54],"analysis":[55],"suggests":[56],"that":[57],"these":[58],"align":[61],"with":[62],"semantic":[63],"distinctions,":[64],"similar":[66,79],"prompts":[67],"cluster":[68],"within":[69],"regions,":[70],"activations":[72],"from":[73],"same":[75],"lead":[77],"next":[80],"token":[81],"predictions.":[82],"This":[83],"work":[84],"provides":[85],"a":[86],"promising":[87],"research":[88],"direction":[89],"for":[90],"understanding":[91],"complexity":[93],"neural":[95],"networks,":[96],"shedding":[97],"light":[98],"on":[99],"dynamics,":[101],"advancing":[103],"interpretability.":[104]},"counts_by_year":[{"year":2026,"cited_by_count":4},{"year":2025,"cited_by_count":1},{"year":2024,"cited_by_count":1}],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2024-10-28T00:00:00"}