{"id":"https://openalex.org/W7202197917","doi":"https://doi.org/10.48550/arxiv.2608.09351","title":"Test-Time Augmentation for LLMs: Where to Inject Diversity at Inference Time","display_name":"Test-Time Augmentation for LLMs: Where to Inject Diversity at Inference Time","publication_year":2026,"publication_date":"2026-08-10","ids":{"openalex":"https://openalex.org/W7202197917","doi":"https://doi.org/10.48550/arxiv.2608.09351"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2608.09351","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.09351","pdf_url":"https://arxiv.org/pdf/2608.09351","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","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2608.09351","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5076673005","display_name":"Nikita Kozodoi","orcid":"https://orcid.org/0000-0002-6186-1109","observed_orcids":["https://orcid.org/0000-0002-6186-1109"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Kozodoi, Nikita","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5129654063","display_name":"Zainab Afolabi","orcid":"https://orcid.org/0009-0004-6968-4460","observed_orcids":["https://orcid.org/0009-0004-6968-4460"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Afolabi, Zainab","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5140773840","display_name":"Jack Butler","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Butler, Jack","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":true,"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/T10028","display_name":"Topic Modeling","score":0.7896000146865845,"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":0.7896000146865845,"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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.0843999981880188,"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":0.0608999989926815,"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/test-time-augmentation","display_name":"test-time augmentation","score":0.9990000128746033},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.8029999732971191},{"id":"https://openalex.org/keywords/response-aggregation","display_name":"response aggregation","score":0.5870000123977661},{"id":"https://openalex.org/keywords/input-disturbances","display_name":"input perturbation","score":0.6460000276565552},{"id":"https://openalex.org/keywords/paraphrasing","display_name":"paraphrase","score":0.8040000200271606},{"id":"https://openalex.org/keywords/self-consistency","display_name":"self-consistency","score":0.8880000114440918},{"id":"https://openalex.org/keywords/inference-cost","display_name":"inference cost","score":0.39100000262260437},{"id":"https://openalex.org/keywords/claude-haiku","display_name":"Claude Haiku","score":0.6660000085830688}],"concepts":[{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6761999726295471},{"id":"https://openalex.org/C176217482","wikidata":"https://www.wikidata.org/wiki/Q860554","display_name":"Metric (unit)","level":2,"score":0.675000011920929},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6244000196456909},{"id":"https://openalex.org/C2780586882","wikidata":"https://www.wikidata.org/wiki/Q7520643","display_name":"Simple (philosophy)","level":2,"score":0.5267000198364258},{"id":"https://openalex.org/C2777735758","wikidata":"https://www.wikidata.org/wiki/Q817765","display_name":"Path (computing)","level":2,"score":0.5052000284194946},{"id":"https://openalex.org/C42058472","wikidata":"https://www.wikidata.org/wiki/Q810214","display_name":"Base (topology)","level":2,"score":0.5016999840736389},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4925000071525574},{"id":"https://openalex.org/C99844830","wikidata":"https://www.wikidata.org/wiki/Q102441924","display_name":"Scaling","level":2,"score":0.4893999993801117},{"id":"https://openalex.org/C140779682","wikidata":"https://www.wikidata.org/wiki/Q210868","display_name":"Sampling (signal processing)","level":3,"score":0.47940000891685486},{"id":"https://openalex.org/C2781316041","wikidata":"https://www.wikidata.org/wiki/Q1230584","display_name":"Diversity (politics)","level":2,"score":0.46889999508857727},{"id":"https://openalex.org/C111335779","wikidata":"https://www.wikidata.org/wiki/Q3454686","display_name":"Reduction (mathematics)","level":2,"score":0.41370001435279846},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.37929999828338623},{"id":"https://openalex.org/C90329073","wikidata":"https://www.wikidata.org/wiki/Q914232","display_name":"Ask price","level":2,"score":0.36739999055862427},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3497999906539917},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.30880001187324524},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.29989999532699585},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.289900004863739},{"id":"https://openalex.org/C2780898871","wikidata":"https://www.wikidata.org/wiki/Q860554","display_name":"Performance metric","level":2,"score":0.2824999988079071},{"id":"https://openalex.org/C148043351","wikidata":"https://www.wikidata.org/wiki/Q4456944","display_name":"Current (fluid)","level":2,"score":0.2773999869823456},{"id":"https://openalex.org/C61797465","wikidata":"https://www.wikidata.org/wiki/Q1188986","display_name":"Term (time)","level":2,"score":0.2558000087738037},{"id":"https://openalex.org/C91682802","wikidata":"https://www.wikidata.org/wiki/Q620538","display_name":"Multidimensional scaling","level":2,"score":0.2502000033855438}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2608.09351","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.09351","pdf_url":"https://arxiv.org/pdf/2608.09351","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.2608.09351","doi":"https://doi.org/10.48550/arxiv.2608.09351","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2608.09351","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2608.09351","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.09351","pdf_url":"https://arxiv.org/pdf/2608.09351","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":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7202197917.pdf","grobid_xml":"https://content.openalex.org/works/W7202197917.grobid-xml"},"referenced_works_count":34,"referenced_works":["https://openalex.org/W3083410900","https://openalex.org/W3202033766","https://openalex.org/W4283728557","https://openalex.org/W4285077564","https://openalex.org/W4365205411","https://openalex.org/W4365563505","https://openalex.org/W4387635374","https://openalex.org/W4388514997","https://openalex.org/W4389519849","https://openalex.org/W4391833261","https://openalex.org/W4393161007","https://openalex.org/W4400526036","https://openalex.org/W4401043133","https://openalex.org/W4402671620","https://openalex.org/W4402716477","https://openalex.org/W4403365357","https://openalex.org/W4404392017","https://openalex.org/W4404781507","https://openalex.org/W4404792934","https://openalex.org/W4405784016","https://openalex.org/W4406058761","https://openalex.org/W4406492615","https://openalex.org/W4406840771","https://openalex.org/W4409628941","https://openalex.org/W4412887774","https://openalex.org/W4412887934","https://openalex.org/W4414817263","https://openalex.org/W4415795202","https://openalex.org/W4416033906","https://openalex.org/W4416621724","https://openalex.org/W4416791639","https://openalex.org/W4417215409","https://openalex.org/W6947892093","https://openalex.org/W6966427115"],"related_works":["https://openalex.org/W4416260007","https://openalex.org/W4407690018","https://openalex.org/W4406492615","https://openalex.org/W7201847851","https://openalex.org/W4402671120","https://openalex.org/W7092291745","https://openalex.org/W4417113945","https://openalex.org/W4415279622","https://openalex.org/W4416132299","https://openalex.org/W4417228524"],"abstract_inverted_index":{"Aggregating":[0],"several":[1,45],"LLM":[2],"responses":[3,21,157],"to":[4,85],"the":[5,20,33,48,51,56,62,80,109,125,128,143,150,156,177,189,236],"same":[6],"question":[7,52,223],"can":[8,53,58,65,196],"improve":[9],"accuracy,":[10,168],"but":[11],"every":[12],"method":[13],"that":[14,107,187,194,200,205,212],"does":[15],"so":[16],"must":[17],"decide":[18],"where":[19,87,108],"should":[22],"differ.":[23],"Established":[24],"test-time":[25],"scaling":[26],"methods":[27],"such":[28,216],"as":[29,217],"self-consistency":[30],"vary":[31],"only":[32],"reasoning":[34,237],"path,":[35],"resampling":[36],"from":[37,112],"a":[38,99,135,222,227],"fixed":[39],"input.":[40],"The":[41],"input":[42,126],"modification":[43],"offers":[44],"further":[46],"options:":[47],"wording":[49],"of":[50,79,102,118,124,230],"be":[54,59,66],"rephrased,":[55],"spelling":[57],"perturbed,":[60],"and":[61,82,138,169],"accompanying":[63],"images":[64],"transformed.":[67],"We":[68],"study":[69],"Test-Time":[70],"Augmentation":[71],"(TTA),":[72],"which":[73],"aggregates":[74],"predictions":[75],"across":[76,95],"transformed":[77],"versions":[78],"input,":[81],"use":[83,229],"it":[84,119,202],"ask":[86],"diversity":[88,110,182],"is":[89,183,206,224,226,242],"best":[90],"injected.":[91],"Comparing":[92],"these":[93],"options":[94],"six":[96],"benchmarks":[97],"at":[98,173,244],"matched":[100],"number":[101],"inference":[103,232],"calls,":[104],"we":[105],"find":[106],"comes":[111],"matters":[113],"more":[114,160],"than":[115,234],"how":[116,221],"much":[117],"there":[120],"is.":[121],"Meaning-preserving":[122],"paraphrases":[123],"yield":[127],"largest":[129],"accuracy":[130,141],"gains":[131],"(+1.80":[132],"pp":[133],"over":[134],"single":[136],"call)":[137],"reach":[139],"peak":[140],"with":[142],"fewest":[144],"samples.":[145],"Yet,":[146],"they":[147],"are":[148],"not":[149],"most":[151],"diverse:":[152],"character-level":[153],"noise":[154],"makes":[155],"disagree":[158],"far":[159],"often":[161],"(37.8%":[162],"versus":[163],"27.4%)":[164],"while":[165,198],"gaining":[166],"less":[167],"perturbing":[170],"two":[171],"modalities":[172],"once":[174],"falls":[175],"below":[176],"single-call":[178],"baseline.":[179],"Not":[180],"all":[181],"equally":[184],"useful:":[185],"perturbations":[186],"keep":[188],"meaning":[190],"intact":[191],"produce":[192,203],"disagreement":[193,204],"aggregation":[195],"use,":[197],"those":[199],"corrupt":[201],"wrong.":[207],"Overall,":[208],"our":[209],"results":[210],"suggest":[211],"for":[213],"mid-tier":[214],"LLMs":[215],"Claude":[218],"Haiku,":[219],"varying":[220,235],"phrased":[225],"better":[228],"an":[231],"call":[233],"path.":[238],"Our":[239],"TTA":[240],"implementation":[241],"available":[243],"github.com/aws-samples/sample-genai-reflection-for-bedrock.":[245]},"counts_by_year":[],"updated_date":"2026-11-08T12:12:28.789561","created_date":"2026-08-12T00:00:00"}