{"id":"https://openalex.org/W7164915546","doi":"https://doi.org/10.48550/arxiv.2606.16193","title":"Cascaded Sparse Autoencoders Learn Multi-Level Visual Concepts in Multimodal LLMs","display_name":"Cascaded Sparse Autoencoders Learn Multi-Level Visual Concepts in Multimodal LLMs","publication_year":2026,"publication_date":"2026-06-15","ids":{"openalex":"https://openalex.org/W7164915546","doi":"https://doi.org/10.48550/arxiv.2606.16193"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.16193","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.16193","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":null,"license_id":null,"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.2606.16193","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5137693256","display_name":"Yusong Zhao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhao, Yusong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138754947","display_name":"Hengyi Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Hengyi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138691239","display_name":"Tanuja Ganu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ganu, Tanuja","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138748732","display_name":"Akshay Nambi","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Nambi, Akshay","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5138709092","display_name":"Hao Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Hao","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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.8355000019073486,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.8355000019073486,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.07530000060796738,"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/T10775","display_name":"Generative Adversarial Networks and Image Synthesis","score":0.031300000846385956,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"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/interpretability","display_name":"Interpretability","score":0.8823999762535095},{"id":"https://openalex.org/keywords/feature","display_name":"Feature (linguistics)","score":0.6158000230789185},{"id":"https://openalex.org/keywords/scalability","display_name":"Scalability","score":0.5133000016212463},{"id":"https://openalex.org/keywords/coherence","display_name":"Coherence (philosophical gambling strategy)","score":0.36739999055862427},{"id":"https://openalex.org/keywords/pattern-recognition","display_name":"Pattern recognition (psychology)","score":0.3596000075340271},{"id":"https://openalex.org/keywords/autoencoder","display_name":"Autoencoder","score":0.3506999909877777},{"id":"https://openalex.org/keywords/visualization","display_name":"Visualization","score":0.31679999828338623},{"id":"https://openalex.org/keywords/mirroring","display_name":"Mirroring","score":0.30959999561309814}],"concepts":[{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.8823999762535095},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7228000164031982},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.6711999773979187},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.6158000230789185},{"id":"https://openalex.org/C48044578","wikidata":"https://www.wikidata.org/wiki/Q727490","display_name":"Scalability","level":2,"score":0.5133000016212463},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.4560999870300293},{"id":"https://openalex.org/C2781181686","wikidata":"https://www.wikidata.org/wiki/Q4226068","display_name":"Coherence (philosophical gambling strategy)","level":2,"score":0.36739999055862427},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.3596000075340271},{"id":"https://openalex.org/C101738243","wikidata":"https://www.wikidata.org/wiki/Q786435","display_name":"Autoencoder","level":3,"score":0.3506999909877777},{"id":"https://openalex.org/C36464697","wikidata":"https://www.wikidata.org/wiki/Q451553","display_name":"Visualization","level":2,"score":0.31679999828338623},{"id":"https://openalex.org/C189645446","wikidata":"https://www.wikidata.org/wiki/Q350865","display_name":"Mirroring","level":2,"score":0.30959999561309814},{"id":"https://openalex.org/C59404180","wikidata":"https://www.wikidata.org/wiki/Q17013334","display_name":"Feature learning","level":2,"score":0.3095000088214874},{"id":"https://openalex.org/C33347731","wikidata":"https://www.wikidata.org/wiki/Q285210","display_name":"Stacking","level":2,"score":0.302700012922287},{"id":"https://openalex.org/C2780598303","wikidata":"https://www.wikidata.org/wiki/Q65921492","display_name":"Flexibility (engineering)","level":2,"score":0.29829999804496765},{"id":"https://openalex.org/C77637269","wikidata":"https://www.wikidata.org/wiki/Q7002051","display_name":"Neural coding","level":2,"score":0.2890999913215637},{"id":"https://openalex.org/C52622490","wikidata":"https://www.wikidata.org/wiki/Q1026626","display_name":"Feature extraction","level":2,"score":0.2667999863624573},{"id":"https://openalex.org/C83665646","wikidata":"https://www.wikidata.org/wiki/Q42139305","display_name":"Feature vector","level":2,"score":0.2599000036716461},{"id":"https://openalex.org/C56372850","wikidata":"https://www.wikidata.org/wiki/Q1050404","display_name":"Sparse matrix","level":3,"score":0.2531999945640564},{"id":"https://openalex.org/C22367795","wikidata":"https://www.wikidata.org/wiki/Q7625208","display_name":"Structured prediction","level":2,"score":0.2515000104904175}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.16193","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.16193","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2606.16193","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.16193","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[{"display_name":"Quality Education","score":0.7718725800514221,"id":"https://metadata.un.org/sdg/4"}],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Multimodal":[0],"Large":[1],"Language":[2],"Models":[3],"(MLLMs)":[4],"have":[5],"demonstrated":[6],"strong":[7],"performance":[8],"on":[9,86,137,157],"vision-language":[10],"tasks,":[11],"yet":[12],"their":[13],"internal":[14],"visual":[15,67,139],"representations":[16],"remain":[17],"difficult":[18],"to":[19,28,108],"interpret.":[20],"Sparse":[21],"Autoencoders":[22],"(SAEs)":[23],"provide":[24],"a":[25,82],"scalable":[26],"way":[27],"decompose":[29],"dense":[30],"model":[31],"activations":[32],"into":[33],"sparse,":[34],"interpretable":[35],"features.":[36],"However,":[37],"existing":[38],"SAE":[39,76,84,154],"architectures":[40],"primarily":[41],"recover":[42],"flat":[43],"feature":[44,97],"dictionaries":[45],"and":[46,124,135],"are":[47],"less":[48],"suited":[49],"for":[50,64,101],"explicit":[51],"multi-level":[52],"concept":[53,150,158,165],"organization.":[54],"In":[55],"this":[56],"paper,":[57],"we":[58],"introduce":[59],"cascaded":[60],"sparse":[61,77],"autoencoders":[62],"(CSAEs)":[63],"learning":[65],"hierarchical":[66,149],"concepts":[68],"in":[69,146,171],"MLLMs.":[70],"Rather":[71],"than":[72],"nesting":[73],"or":[74],"stacking":[75],"activation":[78],"codes,":[79],"CSAEs":[80,107,143],"train":[81],"second-level":[83],"directly":[85],"the":[87,91,117,125,163],"decoder":[88],"weights":[89],"of":[90,111,120,127,148],"first-level":[92],"SAE,":[93],"treating":[94],"learned":[95,164],"low-level":[96],"directions":[98],"as":[99],"inputs":[100],"higher-level":[102],"abstraction.":[103],"This":[104],"design":[105],"enables":[106],"learn":[109],"\"concepts":[110],"concepts\"":[112],"while":[113],"avoiding":[114],"drawbacks":[115],"from":[116],"shared-prefix":[118],"coupling":[119],"nesting,":[121],"Matryoshka-style":[122],"hierarchies":[123],"bottlenecks":[126],"naively":[128],"stacked":[129],"SAEs.":[130],"Experiments":[131],"across":[132],"Qwen3-VL,":[133],"Gemma-3,":[134],"LLaVA":[136],"multiple":[138],"datasets":[140],"show":[141],"that":[142,162],"improve":[144],"interpretability":[145],"terms":[147],"coherence":[151],"over":[152],"state-of-the-art":[153],"baselines.":[155],"Results":[156],"steering":[159],"further":[160],"demonstrate":[161],"groups":[166],"support":[167],"effective":[168],"group-level":[169],"interventions":[170],"MLLM":[172],"outputs.":[173]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-06-17T00:00:00"}