{"id":"https://openalex.org/W7166867525","doi":"https://doi.org/10.18653/v1/2026.acl-long.540","title":"KV-Embedding: Training-free Text Embedding via Internal KV Re-routing in Decoder-only LLMs","display_name":"KV-Embedding: Training-free Text Embedding via Internal KV Re-routing in Decoder-only LLMs","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166867525","doi":"https://doi.org/10.18653/v1/2026.acl-long.540"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2026.acl-long.540","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.540","pdf_url":"https://aclanthology.org/2026.acl-long.540.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2026.acl-long.540.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5139786870","display_name":"Yixuan Tang","orcid":null},"institutions":[{"id":"https://openalex.org/I200769079","display_name":"Hong Kong University of Science and Technology","ror":"https://ror.org/00q4vv597","country_code":"HK","type":"education","lineage":["https://openalex.org/I200769079"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Yixuan Tang","raw_affiliation_strings":["The Hong Kong University of Science and Technology"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"The Hong Kong University of Science and Technology","institution_ids":["https://openalex.org/I200769079"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5139715533","display_name":"Yi Yang","orcid":null},"institutions":[{"id":"https://openalex.org/I200769079","display_name":"Hong Kong University of Science and Technology","ror":"https://ror.org/00q4vv597","country_code":"HK","type":"education","lineage":["https://openalex.org/I200769079"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Yi Yang","raw_affiliation_strings":["The Hong Kong University of Science and Technology"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"The Hong Kong University of Science and Technology","institution_ids":["https://openalex.org/I200769079"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I200769079"],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.8563411,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"11773","last_page":"11794"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11273","display_name":"Advanced Graph Neural Networks","score":0.28790000081062317,"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/T11273","display_name":"Advanced Graph Neural Networks","score":0.28790000081062317,"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":0.17020000517368317,"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/T11550","display_name":"Text and Document Classification Technologies","score":0.08590000122785568,"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/noise","display_name":"Noise (video)","score":0.33079999685287476},{"id":"https://openalex.org/keywords/embedding","display_name":"Embedding","score":0.3003000020980835},{"id":"https://openalex.org/keywords/process","display_name":"Process (computing)","score":0.25220000743865967}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.4952000081539154},{"id":"https://openalex.org/C121332964","wikidata":"https://www.wikidata.org/wiki/Q413","display_name":"Physics","level":0,"score":0.3522000014781952},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.33079999685287476},{"id":"https://openalex.org/C41608201","wikidata":"https://www.wikidata.org/wiki/Q980509","display_name":"Embedding","level":2,"score":0.3003000020980835},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3003000020980835},{"id":"https://openalex.org/C127413603","wikidata":"https://www.wikidata.org/wiki/Q11023","display_name":"Engineering","level":0,"score":0.2603999972343445},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.25220000743865967},{"id":"https://openalex.org/C119599485","wikidata":"https://www.wikidata.org/wiki/Q43035","display_name":"Electrical engineering","level":1,"score":0.2515999972820282},{"id":"https://openalex.org/C28490314","wikidata":"https://www.wikidata.org/wiki/Q189436","display_name":"Speech recognition","level":1,"score":0.24719999730587006},{"id":"https://openalex.org/C76155785","wikidata":"https://www.wikidata.org/wiki/Q418","display_name":"Telecommunications","level":1,"score":0.2281000018119812}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.acl-long.540","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.540","pdf_url":"https://aclanthology.org/2026.acl-long.540.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.acl-long.540","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.540","pdf_url":"https://aclanthology.org/2026.acl-long.540.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166867525.pdf","grobid_xml":"https://content.openalex.org/works/W7166867525.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"While":[0],"LLMs":[1],"are":[2],"powerful":[3],"embedding":[4],"backbones,":[5],"their":[6],"application":[7],"in":[8],"training-free":[9,124],"settings":[10],"faces":[11],"two":[12],"structural":[13],"challenges:":[14],"causal":[15],"attention":[16],"restricts":[17],"early":[18],"tokens":[19,87],"from":[20],"accessing":[21],"subsequent":[22],"context,":[23],"and":[24,116,153],"the":[25,47,56,59,64,75],"next-token":[26],"prediction":[27],"objective":[28],"biases":[29],"representations":[30],"toward":[31],"generation":[32],"rather":[33],"than":[34],"semantic":[35],"compression.To":[36],"address":[37],"these":[38,78],"limitations,":[39],"we":[40,84,100,154],"propose":[41],"KV-Embedding,":[42],"a":[43,71,81,93],"framework":[44],"that":[45,58,120,142],"activates":[46],"latent":[48],"representation":[49,165],"power":[50],"of":[51,63,74,161],"frozen":[52],"LLMs.Our":[53],"method":[54],"leverages":[55],"observation":[57],"key-value":[60],"(KV)":[61],"states":[62,79],"final":[65],"token":[66],"at":[67],"each":[68],"layer":[69,104],"encode":[70],"compressed":[72],"view":[73],"sequence.By":[76],"re-routing":[77],"as":[80],"prepended":[82],"prefix,":[83],"enable":[85],"all":[86],"to":[88,128,137,150],"access":[89],"sequence-level":[90],"context":[91],"within":[92],"single":[94],"forward":[95],"pass.To":[96],"ensure":[97],"model-agnostic":[98],"applicability,":[99],"introduce":[101],"an":[102,147],"automated":[103],"selection":[105],"strategy":[106],"based":[107],"on":[108,111,134],"intrinsic":[109],"dimensionality.Evaluations":[110],"MTEB":[112],"across":[113],"Qwen,":[114],"Mistral,":[115],"Llama":[117],"backbones":[118],"show":[119],"KV-Embedding":[121],"outperforms":[122],"existing":[123],"baselines":[125],"by":[126],"up":[127,136],"10%,":[129],"while":[130],"maintaining":[131],"robust":[132],"performance":[133],"sequences":[135],"4,096":[138],"tokens.These":[139],"results":[140],"demonstrate":[141],"internal":[143],"state":[144],"manipulation":[145],"offers":[146],"efficient":[148],"alternative":[149],"input":[151],"modification,":[152],"hope":[155],"this":[156],"work":[157],"encourages":[158],"further":[159],"exploration":[160],"LLM":[162],"internals":[163],"for":[164],"learning.":[166]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2026-07-02T00:00:00"}