{"id":"https://openalex.org/W4416036762","doi":"https://doi.org/10.18653/v1/2025.emnlp-main.578","title":"Reward-Shifted Speculative Sampling Is An Efficient Test-Time Weak-to-Strong Aligner","display_name":"Reward-Shifted Speculative Sampling Is An Efficient Test-Time Weak-to-Strong Aligner","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4416036762","doi":"https://doi.org/10.18653/v1/2025.emnlp-main.578"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2025.emnlp-main.578","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.578","pdf_url":"https://aclanthology.org/2025.emnlp-main.578.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 2025 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},"type":"article","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2025.emnlp-main.578.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5008288503","display_name":"Bolian Li","orcid":"https://orcid.org/0000-0002-1977-0764"},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Bolian Li","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5105683309","display_name":"Yanran Wu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yanran Wu","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5075584312","display_name":"X. L. Luo","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xinyu Luo","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5053555212","display_name":"Rihong Zhang","orcid":"https://orcid.org/0000-0002-2461-9233"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ruqi Zhang","raw_affiliation_strings":[],"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":4,"corresponding_author_ids":["https://openalex.org/A5008288503"],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.31724285,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"11479","last_page":"11489"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10136","display_name":"Statistical Methods and Inference","score":0.19589999318122864,"subfield":{"id":"https://openalex.org/subfields/2613","display_name":"Statistics and Probability"},"field":{"id":"https://openalex.org/fields/26","display_name":"Mathematics"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T10136","display_name":"Statistical Methods and Inference","score":0.19589999318122864,"subfield":{"id":"https://openalex.org/subfields/2613","display_name":"Statistics and Probability"},"field":{"id":"https://openalex.org/fields/26","display_name":"Mathematics"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T12879","display_name":"Distributed Sensor Networks and Detection Algorithms","score":0.07190000265836716,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"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/T10500","display_name":"Sparse and Compressive Sensing Techniques","score":0.06560000032186508,"subfield":{"id":"https://openalex.org/subfields/2206","display_name":"Computational Mechanics"},"field":{"id":"https://openalex.org/fields/22","display_name":"Engineering"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/sampling","display_name":"Sampling (signal processing)","score":0.42640000581741333},{"id":"https://openalex.org/keywords/noise","display_name":"Noise (video)","score":0.2784999907016754},{"id":"https://openalex.org/keywords/aliasing","display_name":"Aliasing","score":0.24480000138282776},{"id":"https://openalex.org/keywords/computation","display_name":"Computation","score":0.23100000619888306}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5917999744415283},{"id":"https://openalex.org/C140779682","wikidata":"https://www.wikidata.org/wiki/Q210868","display_name":"Sampling (signal processing)","level":3,"score":0.42640000581741333},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.40389999747276306},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.2784999907016754},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.2768000066280365},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.26159998774528503},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.25619998574256897},{"id":"https://openalex.org/C4069607","wikidata":"https://www.wikidata.org/wiki/Q868732","display_name":"Aliasing","level":3,"score":0.24480000138282776},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.23100000619888306},{"id":"https://openalex.org/C2778755073","wikidata":"https://www.wikidata.org/wiki/Q10858537","display_name":"Scale (ratio)","level":2,"score":0.22689999639987946}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2025.emnlp-main.578","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.578","pdf_url":"https://aclanthology.org/2025.emnlp-main.578.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 2025 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2025.emnlp-main.578","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.emnlp-main.578","pdf_url":"https://aclanthology.org/2025.emnlp-main.578.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 2025 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4416036762.pdf","grobid_xml":"https://content.openalex.org/works/W4416036762.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Aligning":[0],"large":[1],"language":[2],"models":[3],"(LLMs)":[4],"with":[5,90],"human":[6,91],"preferences":[7],"has":[8,17],"become":[9],"a":[10,59,144],"critical":[11],"step":[12],"in":[13,83,149],"their":[14,47],"development.Recent":[15],"research":[16],"increasingly":[18],"focused":[19],"on":[20],"test-time":[21,38,74,150],"alignment,":[22],"where":[23],"additional":[24],"compute":[25],"is":[26,88],"allocated":[27],"during":[28],"inference":[29,44,147],"to":[30,63,68,118],"enhance":[31],"LLM":[32],"safety":[33],"and":[34,110,133,159],"reasoning":[35],"capabilities.However,":[36],"these":[37],"alignment":[39,152],"techniques":[40],"often":[41],"incur":[42],"substantial":[43],"costs,":[45],"limiting":[46],"practical":[48],"application.We":[49],"are":[50],"inspired":[51],"by":[52,128],"the":[53,70,77,85,94,102,106,111,120,130],"speculative":[54],"sampling":[55],"acceleration,":[56],"which":[57,84],"leverages":[58],"small":[60],"draft":[61,86,108],"model":[62,87,96,109,114],"efficiently":[64],"predict":[65],"future":[66],"tokens,":[67],"address":[69],"efficiency":[71],"bottleneck":[72],"of":[73],"alignment.We":[75],"introduce":[76],"reward-Shifted":[78],"Speculative":[79],"Sampling":[80],"(SSS)":[81],"algorithm,":[82],"aligned":[89,107],"preferences,":[92],"while":[93],"target":[95,113],"remains":[97],"unchanged.We":[98],"theoretically":[99],"demonstrate":[100],"that":[101],"distributional":[103],"shift":[104],"between":[105],"unaligned":[112],"can":[115],"be":[116],"exploited":[117],"recover":[119],"RLHF":[121],"optimal":[122],"solution":[123],"without":[124],"actually":[125],"obtaining":[126],"it,":[127],"modifying":[129],"acceptance":[131],"criterion":[132],"bonus":[134],"token":[135],"distribution.Our":[136],"algorithm":[137],"achieves":[138],"superior":[139],"gold":[140],"reward":[141],"scores":[142],"at":[143],"significantly":[145],"reduced":[146],"cost":[148],"weak-to-strong":[151],"experiments,":[153],"thereby":[154],"validating":[155],"both":[156],"its":[157],"effectiveness":[158],"efficiency.":[160],"1":[161]},"counts_by_year":[],"updated_date":"2026-03-12T06:13:28.667946","created_date":"2025-11-08T00:00:00"}