{"id":"https://openalex.org/W4407093370","doi":"https://doi.org/10.48550/arxiv.2501.18838","title":"Partially Rewriting a Transformer in Natural Language","display_name":"Partially Rewriting a Transformer in Natural Language","publication_year":2025,"publication_date":"2025-01-31","ids":{"openalex":"https://openalex.org/W4407093370","doi":"https://doi.org/10.48550/arxiv.2501.18838"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2501.18838","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2501.18838","pdf_url":"https://arxiv.org/pdf/2501.18838","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":"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/2501.18838","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5149744284","display_name":"Goncalo Paulo","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Paulo, Gon\u00e7alo","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5088767744","display_name":"Nora Belrose","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Belrose, Nora","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":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":1.0,"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":1.0,"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":0.0,"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":0.0,"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/mechanistic-interpretability","display_name":"mechanistic interpretability","score":0.9779999852180481},{"id":"https://openalex.org/keywords/natural-language-explanations","display_name":"natural language explanations","score":0.7549999952316284},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.7979999780654907},{"id":"https://openalex.org/keywords/transcoder","display_name":"transcoder","score":0.6060000061988831},{"id":"https://openalex.org/keywords/sparse-autoencoders","display_name":"sparse autoencoders","score":0.902999997138977},{"id":"https://openalex.org/keywords/residual-stream","display_name":"residual stream","score":0.9319999814033508},{"id":"https://openalex.org/keywords/ablation","display_name":"ablation","score":0.527999997138977}],"concepts":[{"id":"https://openalex.org/C154690210","wikidata":"https://www.wikidata.org/wiki/Q1668499","display_name":"Rewriting","level":2,"score":0.8027279376983643},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.5252071022987366},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.4517674446105957},{"id":"https://openalex.org/C199360897","wikidata":"https://www.wikidata.org/wiki/Q9143","display_name":"Programming language","level":1,"score":0.4382171928882599},{"id":"https://openalex.org/C127413603","wikidata":"https://www.wikidata.org/wiki/Q11023","display_name":"Engineering","level":0,"score":0.1143261194229126},{"id":"https://openalex.org/C119599485","wikidata":"https://www.wikidata.org/wiki/Q43035","display_name":"Electrical engineering","level":1,"score":0.08460861444473267},{"id":"https://openalex.org/C165801399","wikidata":"https://www.wikidata.org/wiki/Q25428","display_name":"Voltage","level":2,"score":0.0}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2501.18838","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2501.18838","pdf_url":"https://arxiv.org/pdf/2501.18838","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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2501.18838","doi":"https://doi.org/10.48550/arxiv.2501.18838","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2501.18838","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:2501.18838","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2501.18838","pdf_url":"https://arxiv.org/pdf/2501.18838","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":"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":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4407093370.pdf","grobid_xml":"https://content.openalex.org/works/W4407093370.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4399038319","https://openalex.org/W7221930924","https://openalex.org/W4385565237","https://openalex.org/W4387595681","https://openalex.org/W7130549209","https://openalex.org/W4403995319","https://openalex.org/W4296634084","https://openalex.org/W7163069817","https://openalex.org/W3202537109","https://openalex.org/W4411638659"],"abstract_inverted_index":{"The":[0],"greatest":[1],"ambition":[2],"of":[3,50,87,99,160],"mechanistic":[4],"interpretability":[5,73],"is":[6,17,131],"to":[7,20,34,75,114,134,177],"completely":[8],"rewrite":[9,36],"deep":[10],"neural":[11],"networks":[12,53],"in":[13,54,129],"a":[14,37,58,66,153],"format":[15],"that":[16,171],"more":[18,172],"amenable":[19],"human":[21],"understanding,":[22],"while":[23],"preserving":[24],"their":[25],"behavior":[26],"and":[27,69,105,164],"performance.":[28],"In":[29],"this":[30,88,150],"paper,":[31],"we":[32,110],"attempt":[33],"partially":[35],"large":[38],"language":[39,44],"model":[40],"using":[41,152],"simple":[42],"natural":[43],"explanations.":[45],"We":[46,81,145],"first":[47,85],"approximate":[48],"one":[49],"the":[51,55,84,97,106,112,119,126,137,142,147,157,161,182],"feedforward":[52],"LLM":[56],"with":[57,61,91,141],"wider":[59],"MLP":[60,90,139],"sparsely":[62],"activating":[63],"neurons":[64],"-":[65,68],"transcoder":[67],"use":[70],"an":[71,92],"automated":[72],"pipeline":[74],"generate":[76],"explanations":[77,174],"for":[78],"these":[79,116],"neurons.":[80],"then":[82],"replace":[83],"layer":[86,163],"sparse":[89,138,154],"LLM-based":[93],"simulator,":[94],"which":[95,115],"predicts":[96],"activation":[98],"each":[100],"neuron":[101],"given":[102],"its":[103],"explanation":[104],"surrounding":[107],"context.":[108],"Finally,":[109],"measure":[111],"degree":[113],"modifications":[117],"distort":[118],"model's":[120,127],"final":[121],"output.":[122],"With":[123],"our":[124],"pipeline,":[125],"increase":[128],"loss":[130],"statistically":[132],"similar":[133,166],"entirely":[135],"replacing":[136],"output":[140],"zero":[143,183],"vector.":[144],"employ":[146],"same":[148,162],"protocol,":[149],"time":[151],"autoencoder,":[155],"on":[156],"residual":[158],"stream":[159],"obtain":[165],"results.":[167],"These":[168],"results":[169],"suggest":[170],"detailed":[173],"are":[175],"needed":[176],"improve":[178],"performance":[179],"substantially":[180],"above":[181],"ablation":[184],"baseline.":[185]},"counts_by_year":[],"updated_date":"2026-09-25T08:18:29.962427","created_date":"2025-02-04T00:00:00"}