{"id":"https://openalex.org/W4416878728","doi":"https://doi.org/10.48550/arxiv.2504.18574","title":"Understanding the Skill Gap in Recurrent Language Models: The Role of the Gather-and-Aggregate Mechanism","display_name":"Understanding the Skill Gap in Recurrent Language Models: The Role of the Gather-and-Aggregate Mechanism","publication_year":2025,"publication_date":"2025-04-22","ids":{"openalex":"https://openalex.org/W4416878728","doi":"https://doi.org/10.48550/arxiv.2504.18574"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2504.18574","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2504.18574","pdf_url":"https://arxiv.org/pdf/2504.18574","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/2504.18574","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Bick, Aviv","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I74973139","display_name":"Carnegie Mellon University","ror":"https://ror.org/05x2bcf33","country_code":"US","type":"education","lineage":["https://openalex.org/I74973139"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Bick, Aviv","raw_affiliation_strings":["Carnegie Mellon University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Carnegie Mellon University","institution_ids":["https://openalex.org/I74973139"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5009547049","display_name":"Eric P. Xing","orcid":"https://orcid.org/0009-0005-9158-4201","observed_orcids":["https://orcid.org/0009-0005-9158-4201","https://orcid.org/0000-0002-3683-4280"]},"institutions":[{"id":"https://openalex.org/I74973139","display_name":"Carnegie Mellon University","ror":"https://ror.org/05x2bcf33","country_code":"US","type":"education","lineage":["https://openalex.org/I74973139"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Xing, Eric","raw_affiliation_strings":["Carnegie Mellon University MBZUAI"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Carnegie Mellon University MBZUAI","institution_ids":["https://openalex.org/I74973139"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5025386668","display_name":"Albert Gu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I74973139","display_name":"Carnegie Mellon University","ror":"https://ror.org/05x2bcf33","country_code":"US","type":"education","lineage":["https://openalex.org/I74973139"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Gu, Albert","raw_affiliation_strings":["Carnegie Mellon University Cartesia"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Carnegie Mellon University Cartesia","institution_ids":["https://openalex.org/I74973139"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I74973139"],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":1,"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/T10028","display_name":"Topic Modeling","score":0.9057999849319458,"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/T10028","display_name":"Topic Modeling","score":0.9057999849319458,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.046799998730421066,"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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.01489999983459711,"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/state-space-models","display_name":"state-space model","score":0.8579999804496765},{"id":"https://openalex.org/keywords/attention-patterns","display_name":"attention patterns","score":0.4560000002384186},{"id":"https://openalex.org/keywords/hybrid-model","display_name":"hybrid model","score":0.503000020980835},{"id":"https://openalex.org/keywords/long-sequence-modeling","display_name":"long sequence modeling","score":0.5080000162124634},{"id":"https://openalex.org/keywords/llama-3-1","display_name":"Llama 3.1","score":0.7210000157356262}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7519999742507935},{"id":"https://openalex.org/C2780513914","wikidata":"https://www.wikidata.org/wiki/Q18210350","display_name":"Bottleneck","level":2,"score":0.7032999992370605},{"id":"https://openalex.org/C197129107","wikidata":"https://www.wikidata.org/wiki/Q1921621","display_name":"Merge (version control)","level":2,"score":0.6399000287055969},{"id":"https://openalex.org/C143273055","wikidata":"https://www.wikidata.org/wiki/Q2382794","display_name":"Delegate","level":2,"score":0.5309000015258789},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.5228999853134155},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.5156999826431274},{"id":"https://openalex.org/C4679612","wikidata":"https://www.wikidata.org/wiki/Q866298","display_name":"Aggregate (composite)","level":2,"score":0.5144000053405762},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5073000192642212},{"id":"https://openalex.org/C44291984","wikidata":"https://www.wikidata.org/wiki/Q1074173","display_name":"Question answering","level":2,"score":0.484499990940094},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.4117000102996826},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.4065999984741211},{"id":"https://openalex.org/C70437156","wikidata":"https://www.wikidata.org/wiki/Q7228652","display_name":"Pooling","level":2,"score":0.38269999623298645},{"id":"https://openalex.org/C195324797","wikidata":"https://www.wikidata.org/wiki/Q33742","display_name":"Natural language","level":2,"score":0.3693000078201294},{"id":"https://openalex.org/C171018156","wikidata":"https://www.wikidata.org/wiki/Q7370306","display_name":"Rotation formalisms in three dimensions","level":2,"score":0.3352999985218048},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3068000078201294},{"id":"https://openalex.org/C2983448237","wikidata":"https://www.wikidata.org/wiki/Q1078276","display_name":"Language understanding","level":2,"score":0.26910001039505005},{"id":"https://openalex.org/C89611455","wikidata":"https://www.wikidata.org/wiki/Q6804646","display_name":"Mechanism (biology)","level":2,"score":0.26840001344680786},{"id":"https://openalex.org/C48103436","wikidata":"https://www.wikidata.org/wiki/Q599031","display_name":"State (computer science)","level":2,"score":0.2669000029563904},{"id":"https://openalex.org/C2779903281","wikidata":"https://www.wikidata.org/wiki/Q6888026","display_name":"Modalities","level":2,"score":0.2619999945163727},{"id":"https://openalex.org/C77618280","wikidata":"https://www.wikidata.org/wiki/Q1155772","display_name":"Scheme (mathematics)","level":2,"score":0.25440001487731934}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2504.18574","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2504.18574","pdf_url":"https://arxiv.org/pdf/2504.18574","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.2504.18574","doi":"https://doi.org/10.48550/arxiv.2504.18574","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2504.18574","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:2504.18574","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2504.18574","pdf_url":"https://arxiv.org/pdf/2504.18574","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":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4416878728.pdf","grobid_xml":"https://content.openalex.org/works/W4416878728.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7128813462","https://openalex.org/W7133562900","https://openalex.org/W4404782788","https://openalex.org/W4416617107","https://openalex.org/W4404388485","https://openalex.org/W4415328410","https://openalex.org/W4415097162","https://openalex.org/W4394867439","https://openalex.org/W7213413142","https://openalex.org/W7157116623"],"abstract_inverted_index":{"State-space":[0],"models":[1,40],"(SSMs)":[2],"offer":[3],"efficient":[4],"alternatives":[5],"to":[6,115,151,218,243],"Transformers":[7],"for":[8,83,209],"long":[9],"sequences,":[10],"but":[11],"their":[12,220],"fixed-size":[13],"recurrent":[14],"state":[15],"limits":[16],"capability":[17],"on":[18,46,135,143],"algorithmic":[19],"tasks,":[20],"such":[21,157],"as":[22,130,158],"retrieving":[23,103],"past":[24],"context.":[25],"In":[26,70],"this":[27,120],"work,":[28],"we":[29,88],"examine":[30],"how":[31,217],"in-context":[32],"retrieval":[33,121,168,190,260],"operates":[34],"in":[35,75,98,108,171,193,251],"Transformer-":[36],"and":[37,41,161,261],"SSM-based":[38],"language":[39,202],"find":[42,223],"that":[43,90,166,224],"both":[44,71],"rely":[45],"a":[47,52,67,76,92,99,195,206,233,247,252],"similar":[48],"Gather-and-Aggregate":[49],"(G&A)":[50],"mechanism:":[51],"Gather":[53,94],"Head":[54,64,97],"extracts":[55],"relevant":[56],"information":[57],"pieces":[58],"from":[59,113],"context,":[60],"which":[61],"an":[62,256],"Aggregate":[63,96,241],"integrates":[65],"into":[66],"single":[68,93,248],"representation.":[69],"architectures,":[72],"G&A":[73,139,185,249],"concentrates":[74],"few":[77,196,234],"heads,":[78,173,197],"forming":[79],"critical":[80],"bottlenecks":[81],"even":[82],"simple":[84],"retrieval.":[85],"For":[86],"example,":[87],"show":[89,165],"disabling":[91],"or":[95],"pruned":[100,132],"Llama-3.1-8B":[101],"impairs":[102],"the":[104,131,180,188,200,238],"correct":[105],"answer":[106],"letter":[107],"MMLU,":[109],"reducing":[110],"its":[111],"accuracy":[112],"66%":[114],"25%":[116],"(random":[117],"guessing).":[118],"Moreover,":[119],"bottleneck":[122,148],"can":[123],"obscure":[124],"limited":[125],"knowledge":[126,145],"demands":[127],"of":[128,179,240],"tasks":[129,152],"model":[133],"succeeds":[134],"MMLU":[136],"with":[137,232,255],"functioning":[138],"heads":[140],"yet":[141],"fails":[142],"other":[144],"benchmarks.":[146],"The":[147],"similarly":[149],"extends":[150],"where":[153,228],"SSMs":[154,229],"typically":[155],"underperform,":[156],"GSM8K,":[159],"BBH,":[160],"dialogue":[162],"comprehension.":[163],"We":[164,222],"SSMs'":[167],"challenges":[169],"manifest":[170],"these":[172],"creating":[174],"smoother":[175],"attention":[176,235,257],"patterns":[177],"instead":[178],"sharp":[181],"token":[182],"transitions":[183],"effective":[184],"requires.":[186],"Thus,":[187],"Transformer-SSM":[189],"gap":[191,214],"exists":[192],"just":[194],"rather":[198],"than":[199],"entire":[201],"model.":[203],"This":[204],"suggests":[205],"unified":[207],"explanation":[208],"Transformer":[210],"vs.":[211],"SSM":[212,254],"performance":[213],"while":[215],"showing":[216],"merge":[219],"strengths.":[221],"pretrained":[225,253],"hybrid":[226],"models,":[227],"are":[230],"combined":[231],"layers,":[236],"delegate":[237],"role":[239],"Heads":[242],"attention.":[244],"Similarly,":[245],"replacing":[246],"head":[250],"variant":[258],"boosts":[259],"benchmark":[262],"scores.":[263]},"counts_by_year":[{"year":2026,"cited_by_count":1}],"updated_date":"2026-11-04T10:54:05.112534","created_date":"2025-10-10T00:00:00"}