{"id":"https://openalex.org/W7166845131","doi":"https://doi.org/10.18653/v1/2026.findings-acl.50","title":"TARo: Token-level Adaptive Routing for LLM Test-time Alignment","display_name":"TARo: Token-level Adaptive Routing for LLM Test-time Alignment","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166845131","doi":"https://doi.org/10.18653/v1/2026.findings-acl.50"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2026.findings-acl.50","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.50","pdf_url":"https://aclanthology.org/2026.findings-acl.50.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":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2026.findings-acl.50.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5139779756","display_name":"Arushi Rai","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Arushi Rai","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139809209","display_name":"Qiang Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Qiang Zhang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5014103790","display_name":"Hanqing Zeng","orcid":"https://orcid.org/0000-0002-2578-2147"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Hanqing Zeng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139773228","display_name":"Yunkai Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yunkai Zhang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139812244","display_name":"Dipesh Tamboli","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Dipesh Tamboli","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139836425","display_name":"Xiangjun Fan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xiangjun Fan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139740080","display_name":"Zhuokai Zhao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhuokai Zhao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5139808581","display_name":"Lizhu Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lizhu Zhang","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":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.79467681,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"1004","last_page":"1025"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10743","display_name":"Software Testing and Debugging Techniques","score":0.4002000093460083,"subfield":{"id":"https://openalex.org/subfields/1712","display_name":"Software"},"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/T10743","display_name":"Software Testing and Debugging Techniques","score":0.4002000093460083,"subfield":{"id":"https://openalex.org/subfields/1712","display_name":"Software"},"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/T12326","display_name":"Network Packet Processing and Optimization","score":0.07620000094175339,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"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/T10142","display_name":"Formal Methods in Verification","score":0.06480000168085098,"subfield":{"id":"https://openalex.org/subfields/1703","display_name":"Computational Theory and Mathematics"},"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/routing","display_name":"Routing (electronic design automation)","score":0.40689998865127563},{"id":"https://openalex.org/keywords/noise","display_name":"Noise (video)","score":0.2827000021934509},{"id":"https://openalex.org/keywords/class","display_name":"Class (philosophy)","score":0.27410000562667847},{"id":"https://openalex.org/keywords/key","display_name":"Key (lock)","score":0.2603999972343445},{"id":"https://openalex.org/keywords/process","display_name":"Process (computing)","score":0.2590999901294708}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5965999960899353},{"id":"https://openalex.org/C74172769","wikidata":"https://www.wikidata.org/wiki/Q1446839","display_name":"Routing (electronic design automation)","level":2,"score":0.40689998865127563},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.2827000021934509},{"id":"https://openalex.org/C31258907","wikidata":"https://www.wikidata.org/wiki/Q1301371","display_name":"Computer network","level":1,"score":0.2825999855995178},{"id":"https://openalex.org/C2777212361","wikidata":"https://www.wikidata.org/wiki/Q5127848","display_name":"Class (philosophy)","level":2,"score":0.27410000562667847},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.2603999972343445},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.2590999901294708},{"id":"https://openalex.org/C79403827","wikidata":"https://www.wikidata.org/wiki/Q3988","display_name":"Real-time computing","level":1,"score":0.25270000100135803},{"id":"https://openalex.org/C163258240","wikidata":"https://www.wikidata.org/wiki/Q25342","display_name":"Power (physics)","level":2,"score":0.24789999425411224},{"id":"https://openalex.org/C127413603","wikidata":"https://www.wikidata.org/wiki/Q11023","display_name":"Engineering","level":0,"score":0.23469999432563782}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.findings-acl.50","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.50","pdf_url":"https://aclanthology.org/2026.findings-acl.50.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":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.findings-acl.50","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.50","pdf_url":"https://aclanthology.org/2026.findings-acl.50.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":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"sustainable_development_goals":[{"score":0.405943900346756,"id":"https://metadata.un.org/sdg/13","display_name":"Climate action"}],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166845131.pdf","grobid_xml":"https://content.openalex.org/works/W7166845131.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Large":[0],"language":[1],"models":[2,60],"(LLMs)":[3],"exhibit":[4],"strong":[5],"reasoning":[6,51,96,118],"capabilities":[7],"but":[8,24],"typically":[9],"require":[10],"expensive":[11],"post-training":[12],"to":[13,65,86,100,129,140],"reach":[14],"high":[15],"performance.Recent":[16],"test-time":[17,110,135],"alignment":[18,31,111,136],"methods":[19],"offer":[20],"a":[21,73],"lightweight":[22],"alternative,":[23],"have":[25],"been":[26],"explored":[27],"mainly":[28],"for":[29],"preference":[30,138],"rather":[32],"than":[33],"reasoning.To":[34],"bridge":[35],"this":[36],"gap,":[37],"we":[38,56],"propose":[39],"Token-level":[40],"Adaptive":[41],"Routing":[42],"(TARO)":[43],",":[44],"which":[45],"steers":[46],"frozen":[47],"LLMs":[48],"toward":[49],"structured":[50],"entirely":[52],"at":[53],"inference":[54],"time.Specifically,":[55],"first":[57],"train":[58],"reward":[59,84],"on":[61],"step-wise":[62],"mathematical":[63],"traces":[64],"capture":[66],"fine-grained":[67],"logical":[68],"consistency":[69],"signals,":[70],"then":[71],"introduce":[72],"learnable":[74],"token-level":[75,109],"router":[76],"that":[77,92],"automatically":[78],"controls":[79],"the":[80,83,87],"guidance":[81],"of":[82],"model":[85,104],"base":[88,103],"model.Extensive":[89],"experiments":[90],"show":[91],"TARO":[93,124],"significantly":[94],"improves":[95],"performance":[97],"by":[98],"up":[99],"+22.4%":[101],"over":[102,107],"and":[105,120],"+8.4%":[106],"existing":[108],"methods,":[112],"while":[113],"also":[114,125],"boosting":[115],"out-ofdistribution":[116],"clinical":[117],"(MedXpertQA)":[119],"instruction":[121],"following":[122],"(AlpacaEval).Furthermore,":[123],"generalizes":[126],"from":[127,137],"small":[128],"large":[130],"backbones":[131],"without":[132],"retraining,":[133],"extending":[134],"optimization":[139],"robust,":[141],"cross-domain":[142],"reasoning.":[143]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2026-07-02T00:00:00"}