{"id":"https://openalex.org/W7164341804","doi":"https://doi.org/10.48550/arxiv.2606.11198","title":"The Structural Attention Tax: How Retrieval Format Hijacks In-Context Learning Independent of Content","display_name":"The Structural Attention Tax: How Retrieval Format Hijacks In-Context Learning Independent of Content","publication_year":2026,"publication_date":"2026-04-21","ids":{"openalex":"https://openalex.org/W7164341804","doi":"https://doi.org/10.48550/arxiv.2606.11198"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.11198","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.11198","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":"Preprint"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2606.11198","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5138434776","display_name":"Yuqi Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Yuqi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5138435712","display_name":"Di Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Di","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/T10286","display_name":"Information Retrieval and Search Behavior","score":0.6657999753952026,"subfield":{"id":"https://openalex.org/subfields/1710","display_name":"Information Systems"},"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/T10286","display_name":"Information Retrieval and Search Behavior","score":0.6657999753952026,"subfield":{"id":"https://openalex.org/subfields/1710","display_name":"Information Systems"},"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":0.04569999873638153,"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/T11273","display_name":"Advanced Graph Neural Networks","score":0.03709999844431877,"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/relevance","display_name":"Relevance (law)","score":0.5559999942779541},{"id":"https://openalex.org/keywords/graph","display_name":"Graph","score":0.4431000053882599},{"id":"https://openalex.org/keywords/term","display_name":"Term (time)","score":0.4122999906539917},{"id":"https://openalex.org/keywords/security-token","display_name":"Security token","score":0.38280001282691956},{"id":"https://openalex.org/keywords/semantic-similarity","display_name":"Semantic similarity","score":0.3774999976158142},{"id":"https://openalex.org/keywords/encoder","display_name":"Encoder","score":0.34689998626708984},{"id":"https://openalex.org/keywords/semantics","display_name":"Semantics (computer science)","score":0.33000001311302185},{"id":"https://openalex.org/keywords/language-model","display_name":"Language model","score":0.32359999418258667}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7940999865531921},{"id":"https://openalex.org/C158154518","wikidata":"https://www.wikidata.org/wiki/Q7310970","display_name":"Relevance (law)","level":2,"score":0.5559999942779541},{"id":"https://openalex.org/C23123220","wikidata":"https://www.wikidata.org/wiki/Q816826","display_name":"Information retrieval","level":1,"score":0.48100000619888306},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.4593000113964081},{"id":"https://openalex.org/C132525143","wikidata":"https://www.wikidata.org/wiki/Q141488","display_name":"Graph","level":2,"score":0.4431000053882599},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4268999993801117},{"id":"https://openalex.org/C61797465","wikidata":"https://www.wikidata.org/wiki/Q1188986","display_name":"Term (time)","level":2,"score":0.4122999906539917},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.38280001282691956},{"id":"https://openalex.org/C130318100","wikidata":"https://www.wikidata.org/wiki/Q2268914","display_name":"Semantic similarity","level":2,"score":0.3774999976158142},{"id":"https://openalex.org/C118505674","wikidata":"https://www.wikidata.org/wiki/Q42586063","display_name":"Encoder","level":2,"score":0.34689998626708984},{"id":"https://openalex.org/C184337299","wikidata":"https://www.wikidata.org/wiki/Q1437428","display_name":"Semantics (computer science)","level":2,"score":0.33000001311302185},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.32359999418258667},{"id":"https://openalex.org/C2779530757","wikidata":"https://www.wikidata.org/wiki/Q1207505","display_name":"Quality (philosophy)","level":2,"score":0.3179999887943268},{"id":"https://openalex.org/C205606062","wikidata":"https://www.wikidata.org/wiki/Q5249645","display_name":"Decoupling (probability)","level":2,"score":0.3131999969482422},{"id":"https://openalex.org/C177774035","wikidata":"https://www.wikidata.org/wiki/Q1246948","display_name":"Granularity","level":2,"score":0.30630001425743103},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.30480000376701355},{"id":"https://openalex.org/C44291984","wikidata":"https://www.wikidata.org/wiki/Q1074173","display_name":"Question answering","level":2,"score":0.29409998655319214},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.27379998564720154},{"id":"https://openalex.org/C106624574","wikidata":"https://www.wikidata.org/wiki/Q5270387","display_name":"Diagrammatic reasoning","level":2,"score":0.2720000147819519},{"id":"https://openalex.org/C108583219","wikidata":"https://www.wikidata.org/wiki/Q197536","display_name":"Deep learning","level":2,"score":0.2630999982357025},{"id":"https://openalex.org/C146044194","wikidata":"https://www.wikidata.org/wiki/Q5157334","display_name":"Computational photography","level":4,"score":0.2581000030040741},{"id":"https://openalex.org/C1667742","wikidata":"https://www.wikidata.org/wiki/Q10927554","display_name":"Image retrieval","level":3,"score":0.2563999891281128}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.11198","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.11198","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":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2606.11198","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.11198","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":"Preprint"},"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":{"Retrieval-augmented":[0],"generation":[1],"(RAG)":[2],"systems":[3],"inject":[4],"external":[5],"knowledge":[6,43],"to":[7,48,79,116,216],"improve":[8],"LLM":[9],"outputs,":[10],"yet":[11],"the":[12,27,39,85,123,133,209,243],"format":[13,114,219],"of":[14,83,245],"injected":[15],"content":[16],"--":[17,23,81],"distinct":[18],"from":[19,208,212,230],"its":[20],"semantic":[21,100,134],"relevance":[22],"can":[24],"independently":[25],"distort":[26],"model's":[28],"attention":[29,41,59,76,97,118,129,160],"distribution.":[30],"We":[31,91,202],"identify":[32],"and":[33,52,101,120,157,171,227],"formalise":[34],"a":[35,93,107,191,231],"phenomenon":[36],"we":[37,175],"term":[38,125,135],"structural":[40,102,124,235],"tax:":[42],"graph":[44],"(KG)":[45],"triples,":[46],"due":[47],"their":[49],"relational":[50],"delimiters":[51],"repeated":[53],"slot":[54],"patterns,":[55],"capture":[56,161],"2-3x":[57],"more":[58],"per":[60],"token":[61],"than":[62],"semantically":[63],"equivalent":[64],"natural-language":[65],"text":[66],"($\\hat{o}$(KG)":[67],"$\\approx$":[68,72],"0.70":[69],"vs.":[70,188],"$\\hat{o}$(neutral)":[71],"0.25),":[73],"compressing":[74],"demonstration":[75,117],"by":[77,224],"up":[78],"42%":[80],"regardless":[82],"whether":[84,137],"triples":[86],"are":[87],"relevant":[88],"or":[89,140],"noise.":[90],"develop":[92],"formal":[94],"framework":[95],"decomposing":[96],"scores":[98],"into":[99],"components":[103],"(Eq.":[104],"2),":[105],"derive":[106,203],"compression":[108],"bound":[109],"(Proposition":[110],"1)":[111],"connecting":[112],"token-level":[113],"bias":[115],"loss,":[119],"show":[121],"that":[122,177,195,241],"governs":[126,136],"how":[127],"much":[128],"is":[130,222],"diverted":[131],"while":[132,234],"this":[138],"helps":[139],"hurts.":[141],"This":[142],"decoupling":[143],"reveals":[144],"two":[145,166],"orthogonal":[146],"axes":[147],"for":[148],"improving":[149],"retrieval-augmented":[150],"ICL:":[151],"optimising":[152],"retrieval":[153,183],"quality":[154],"(semantic":[155],"axis)":[156],"reducing":[158],"format-driven":[159],"(structural":[162],"axis).":[163],"Empirically,":[164],"across":[165],"model":[167],"families":[168],"(Mistral-7B,":[169],"LLaMA-3-8B)":[170],"three":[172],"QA":[173],"benchmarks,":[174],"observe":[176],"source-task":[178],"alignment":[179],"dominates:":[180],"task-matched":[181],"BM25":[182],"achieves":[184],"58-62%":[185],"on":[186],"HotpotQA":[187],"ConceptNet's":[189],"25-27%,":[190],">30":[192],"pp":[193],"gap":[194],"dwarfs":[196],"all":[197],"gating":[198],"strategies":[199,207],"($\\leq$2":[200],"pp).":[201],"five":[204],"structure-aware":[205],"mitigation":[206],"framework,":[210],"ranging":[211],"zero-cost":[213],"prompt":[214],"modifications":[215],"training-time":[217],"regularisation;":[218],"flattening":[220],"(S3)":[221],"validated":[223],"both":[225],"accuracy":[226],"attention-level":[228],"evidence":[229],"verbalized-triple":[232],"control,":[233],"dispersal":[236],"(S1)":[237],"yields":[238],"mixed":[239],"results":[240],"illuminate":[242],"challenges":[244],"format-level":[246],"intervention.":[247]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-06-12T00:00:00"}