{"id":"https://openalex.org/W7136973508","doi":"https://doi.org/10.48550/arxiv.2603.12875","title":"Test-time RL alignment exposes task familiarity artifacts in LLM benchmarks","display_name":"Test-time RL alignment exposes task familiarity artifacts in LLM benchmarks","publication_year":2026,"publication_date":"2026-03-13","ids":{"openalex":"https://openalex.org/W7136973508","doi":"https://doi.org/10.48550/arxiv.2603.12875"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2603.12875","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.12875","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,"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"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2603.12875","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Wang, Kun","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Kun","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5003606899","display_name":"Reinhard Heckel","orcid":"https://orcid.org/0000-0002-2874-2984"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Heckel, Reinhard","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/T10462","display_name":"Reinforcement Learning in Robotics","score":0.7163000106811523,"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/T10462","display_name":"Reinforcement Learning in Robotics","score":0.7163000106811523,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.06719999760389328,"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":0.03139999881386757,"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/benchmark","display_name":"Benchmark (surveying)","score":0.8287000060081482},{"id":"https://openalex.org/keywords/task","display_name":"Task (project management)","score":0.8187999725341797},{"id":"https://openalex.org/keywords/reinforcement-learning","display_name":"Reinforcement learning","score":0.6251999735832214},{"id":"https://openalex.org/keywords/training-set","display_name":"Training set","score":0.45179998874664307},{"id":"https://openalex.org/keywords/base","display_name":"Base (topology)","score":0.4490000009536743},{"id":"https://openalex.org/keywords/task-analysis","display_name":"Task analysis","score":0.38040000200271606}],"concepts":[{"id":"https://openalex.org/C185798385","wikidata":"https://www.wikidata.org/wiki/Q1161707","display_name":"Benchmark (surveying)","level":2,"score":0.8287000060081482},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.8187999725341797},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7512000203132629},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.6880000233650208},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.6251999735832214},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.5831999778747559},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.45179998874664307},{"id":"https://openalex.org/C42058472","wikidata":"https://www.wikidata.org/wiki/Q810214","display_name":"Base (topology)","level":2,"score":0.4490000009536743},{"id":"https://openalex.org/C175154964","wikidata":"https://www.wikidata.org/wiki/Q380077","display_name":"Task analysis","level":3,"score":0.38040000200271606},{"id":"https://openalex.org/C198531522","wikidata":"https://www.wikidata.org/wiki/Q485146","display_name":"Sample (material)","level":2,"score":0.3734000027179718},{"id":"https://openalex.org/C4554734","wikidata":"https://www.wikidata.org/wiki/Q593744","display_name":"Knowledge base","level":2,"score":0.35249999165534973},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.2761000096797943},{"id":"https://openalex.org/C28006648","wikidata":"https://www.wikidata.org/wiki/Q6934509","display_name":"Multi-task learning","level":3,"score":0.2736999988555908},{"id":"https://openalex.org/C2776145971","wikidata":"https://www.wikidata.org/wiki/Q30673951","display_name":"Labeled data","level":2,"score":0.2563000023365021}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2603.12875","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.12875","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,"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":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2603.12875","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.12875","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,"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"},"sustainable_development_goals":[{"display_name":"Peace, Justice and strong institutions","id":"https://metadata.un.org/sdg/16","score":0.41463178396224976}],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Direct":[0],"evaluation":[1,50,133,147],"of":[2,82,148,189],"LLMs":[3],"on":[4],"benchmarks":[5],"can":[6],"be":[7],"misleading":[8],"because":[9],"comparatively":[10],"strong":[11],"performance":[12,156],"may":[13],"reflect":[14],"task":[15,25,87,190],"familiarity":[16,26],"rather":[17,187],"than":[18],"capability.":[19],"The":[20],"train-before-test":[21],"approach":[22],"controls":[23],"for":[24,70,152],"by":[27],"giving":[28],"each":[29],"model":[30,84,98],"task-relevant":[31],"training":[32,41,120,127],"before":[33],"evaluation,":[34],"originally":[35],"through":[36],"supervised":[37],"finetuning.":[38],"However,":[39],"suitable":[40],"data":[42,55],"is":[43],"often":[44],"hard":[45],"to":[46,85,99],"come":[47],"by,":[48],"and":[49,89,161],"results":[51],"vary":[52],"with":[53,74,93],"the":[54,83,86,97,100,155,177],"chosen.":[56],"In":[57],"this":[58],"paper,":[59],"we":[60,129],"propose":[61],"a":[62,75,79,118,123,144,181],"two-stage":[63],"test-time":[64,91,104],"reinforcement":[65],"learning":[66],"(RL)":[67],"alignment":[68,81,106],"method":[69,107],"train-before-test.":[71],"First,":[72],"RL":[73,92,105],"single":[76],"sample":[77],"provides":[78],"first":[80],"format,":[88],"second,":[90],"majority-voting":[94],"reward":[95],"aligns":[96,108],"benchmark":[101,125],"distribution.":[102],"Our":[103],"similarly":[109],"well":[110],"as":[111],"SFT-based":[112],"train-before":[113],"test,":[114],"but":[115,186],"without":[116,126],"requiring":[117],"task-specific":[119],"set.":[121],"On":[122],"domain-specific":[124],"data,":[128],"show":[130],"that":[131,170],"direct":[132],"underestimates":[134],"base":[135,163],"models":[136,160,164],"which":[137],"perform":[138],"substantially":[139],"better":[140],"once":[141],"aligned,":[142],"yielding":[143],"more":[145],"faithful":[146],"their":[149,162],"capabilities.":[150],"Moreover,":[151],"reasoning":[153,184],"tasks,":[154],"gap":[157],"between":[158],"fine-tuned":[159],"largely":[165],"disappears":[166],"after":[167],"alignment,":[168],"suggesting":[169],"many":[171],"gains":[172],"from":[173],"RLVR/SFT":[174],"reported":[175],"in":[176,183],"literature":[178],"are":[179],"not":[180],"difference":[182],"capability,":[185],"artifacts":[188],"familiarity.":[191]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-03-17T00:00:00"}