{"id":"https://openalex.org/W7139059113","doi":"https://doi.org/10.48550/arxiv.2603.17228","title":"From Drop-off to Recovery: A Mechanistic Analysis of Segmentation in MLLMs","display_name":"From Drop-off to Recovery: A Mechanistic Analysis of Segmentation in MLLMs","publication_year":2026,"publication_date":"2026-03-18","ids":{"openalex":"https://openalex.org/W7139059113","doi":"https://doi.org/10.48550/arxiv.2603.17228"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2603.17228","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.17228","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":false,"raw_source_name":null,"raw_type":"article"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2603.17228","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5129878330","display_name":"Boyong Wu","orcid":null},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Wu, Boyong","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5129988474","display_name":"Sanghwan Kim","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Kim, Sanghwan","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5129962024","display_name":"Zeynep Akata","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Akata, Zeynep","raw_affiliation_strings":[],"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":3,"corresponding_author_ids":["https://openalex.org/A5129878330"],"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":0.792900025844574,"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":0.792900025844574,"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":0.030500000342726707,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.019899999722838402,"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/segmentation","display_name":"Segmentation","score":0.6588000059127808},{"id":"https://openalex.org/keywords/representation","display_name":"Representation (politics)","score":0.5027999877929688},{"id":"https://openalex.org/keywords/image-segmentation","display_name":"Image segmentation","score":0.4754999876022339},{"id":"https://openalex.org/keywords/process","display_name":"Process (computing)","score":0.4648999869823456},{"id":"https://openalex.org/keywords/image","display_name":"Image (mathematics)","score":0.447299987077713},{"id":"https://openalex.org/keywords/pattern-recognition","display_name":"Pattern recognition (psychology)","score":0.42969998717308044},{"id":"https://openalex.org/keywords/security-token","display_name":"Security token","score":0.41909998655319214}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6851000189781189},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.6704000234603882},{"id":"https://openalex.org/C89600930","wikidata":"https://www.wikidata.org/wiki/Q1423946","display_name":"Segmentation","level":2,"score":0.6588000059127808},{"id":"https://openalex.org/C2776359362","wikidata":"https://www.wikidata.org/wiki/Q2145286","display_name":"Representation (politics)","level":3,"score":0.5027999877929688},{"id":"https://openalex.org/C124504099","wikidata":"https://www.wikidata.org/wiki/Q56933","display_name":"Image segmentation","level":3,"score":0.4754999876022339},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.4648999869823456},{"id":"https://openalex.org/C115961682","wikidata":"https://www.wikidata.org/wiki/Q860623","display_name":"Image (mathematics)","level":2,"score":0.447299987077713},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.42969998717308044},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.41909998655319214},{"id":"https://openalex.org/C31972630","wikidata":"https://www.wikidata.org/wiki/Q844240","display_name":"Computer vision","level":1,"score":0.4059000015258789},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.38440001010894775},{"id":"https://openalex.org/C159620131","wikidata":"https://www.wikidata.org/wiki/Q1938983","display_name":"Spatial analysis","level":2,"score":0.32519999146461487},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.29919999837875366},{"id":"https://openalex.org/C9417928","wikidata":"https://www.wikidata.org/wiki/Q1070689","display_name":"Image processing","level":3,"score":0.2757999897003174},{"id":"https://openalex.org/C177284502","wikidata":"https://www.wikidata.org/wiki/Q1005390","display_name":"Adapter (computing)","level":2,"score":0.2655999958515167},{"id":"https://openalex.org/C5339829","wikidata":"https://www.wikidata.org/wiki/Q1425977","display_name":"Machine vision","level":2,"score":0.258899986743927},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.25040000677108765}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2603.17228","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.17228","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":"doi:10.48550/arxiv.2603.17228","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.17228","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":false,"raw_source_name":null,"raw_type":"article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":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],"are":[5],"increasingly":[6],"applied":[7],"to":[8,51],"pixel-level":[9],"vision":[10,37],"tasks,":[11],"yet":[12],"their":[13],"intrinsic":[14],"capacity":[15,25],"for":[16,133],"spatial":[17,70],"understanding":[18],"remains":[19],"poorly":[20],"understood.":[21],"We":[22,42],"investigate":[23],"segmentation":[24,80],"through":[26,88],"a":[27,79,124],"layerwise":[28],"linear":[29],"probing":[30],"evaluation":[31,62],"across":[32],"the":[33,76,99,136],"entire":[34],"MLLM":[35],"pipeline:":[36],"encoder,":[38],"adapter,":[39],"and":[40,60],"LLM.":[41],"further":[43],"conduct":[44],"an":[45,61],"intervention":[46],"based":[47],"attention":[48,55,65,116],"knockout":[49],"analysis":[50,73],"test":[52],"whether":[53],"cross-token":[54],"progressively":[56,86],"refines":[57],"visual":[58,131],"representations,":[59],"of":[63,127,138],"bidirectional":[64,115],"among":[66,117],"image":[67,104,118],"tokens":[68,94,119],"on":[69],"consistency.":[71],"Our":[72],"reveals":[74],"that":[75],"adapter":[77],"introduces":[78],"representation":[81],"drop-off,":[82],"but":[83],"LLM":[84],"layers":[85],"recover":[87],"attention-mediated":[89],"refinement,":[90],"where":[91],"correctly":[92],"classified":[93],"steer":[95],"misclassified":[96],"neighbors":[97],"toward":[98],"correct":[100],"label.":[101],"At":[102],"early":[103],"token":[105],"positions,":[106],"this":[107],"recovery":[108],"is":[109],"bounded":[110],"by":[111],"causal":[112],"attention,":[113],"which":[114],"alleviates.":[120],"These":[121],"findings":[122],"provide":[123],"mechanistic":[125],"account":[126],"how":[128],"MLLMs":[129],"process":[130],"information":[132],"segmentation,":[134],"informing":[135],"design":[137],"future":[139],"segmentation-capable":[140],"models.":[141]},"counts_by_year":[],"updated_date":"2026-03-20T20:54:20.808490","created_date":"2026-03-20T00:00:00"}