{"id":"https://openalex.org/W7154702287","doi":"https://doi.org/10.48550/arxiv.2604.14853","title":"Adaptive Test-Time Compute Allocation for Reasoning LLMs via Constrained Policy Optimization","display_name":"Adaptive Test-Time Compute Allocation for Reasoning LLMs via Constrained Policy Optimization","publication_year":2026,"publication_date":"2026-04-16","ids":{"openalex":"https://openalex.org/W7154702287","doi":"https://doi.org/10.48550/arxiv.2604.14853"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2604.14853","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.14853","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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.2604.14853","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5133905984","display_name":"Zhiyuan Zhai","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhai, Zhiyuan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5057208641","display_name":"Bingcong Li","orcid":"https://orcid.org/0000-0003-1958-4168"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Bingcong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133893337","display_name":"Bingnan Xiao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xiao, Bingnan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133839423","display_name":"Ming Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Ming","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5133905305","display_name":"Xin Eric Wang","orcid":"https://orcid.org/0009-0005-0216-8661"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Xin","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/T10181","display_name":"Natural Language Processing Techniques","score":0.4359000027179718,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.4359000027179718,"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":0.14839999377727509,"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":0.09950000047683716,"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/oracle","display_name":"Oracle","score":0.5939000248908997},{"id":"https://openalex.org/keywords/budget-constraint","display_name":"Budget constraint","score":0.582099974155426},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.5347999930381775},{"id":"https://openalex.org/keywords/lagrangian-relaxation","display_name":"Lagrangian relaxation","score":0.47609999775886536},{"id":"https://openalex.org/keywords/estimator","display_name":"Estimator","score":0.46380001306533813},{"id":"https://openalex.org/keywords/regret","display_name":"Regret","score":0.4602999985218048},{"id":"https://openalex.org/keywords/boosting","display_name":"Boosting (machine learning)","score":0.4083999991416931},{"id":"https://openalex.org/keywords/monotone-polygon","display_name":"Monotone polygon","score":0.36489999294281006},{"id":"https://openalex.org/keywords/heuristics","display_name":"Heuristics","score":0.35109999775886536}],"concepts":[{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.598800003528595},{"id":"https://openalex.org/C55166926","wikidata":"https://www.wikidata.org/wiki/Q2892946","display_name":"Oracle","level":2,"score":0.5939000248908997},{"id":"https://openalex.org/C8505890","wikidata":"https://www.wikidata.org/wiki/Q605095","display_name":"Budget constraint","level":2,"score":0.582099974155426},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5411999821662903},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5347999930381775},{"id":"https://openalex.org/C91765299","wikidata":"https://www.wikidata.org/wiki/Q3424292","display_name":"Lagrangian relaxation","level":2,"score":0.47609999775886536},{"id":"https://openalex.org/C185429906","wikidata":"https://www.wikidata.org/wiki/Q1130160","display_name":"Estimator","level":2,"score":0.46380001306533813},{"id":"https://openalex.org/C50817715","wikidata":"https://www.wikidata.org/wiki/Q79895177","display_name":"Regret","level":2,"score":0.4602999985218048},{"id":"https://openalex.org/C46686674","wikidata":"https://www.wikidata.org/wiki/Q466303","display_name":"Boosting (machine learning)","level":2,"score":0.4083999991416931},{"id":"https://openalex.org/C2834757","wikidata":"https://www.wikidata.org/wiki/Q4925424","display_name":"Monotone polygon","level":2,"score":0.36489999294281006},{"id":"https://openalex.org/C127705205","wikidata":"https://www.wikidata.org/wiki/Q5748245","display_name":"Heuristics","level":2,"score":0.35109999775886536},{"id":"https://openalex.org/C34388435","wikidata":"https://www.wikidata.org/wiki/Q2267362","display_name":"Bounded function","level":2,"score":0.34599998593330383},{"id":"https://openalex.org/C173801870","wikidata":"https://www.wikidata.org/wiki/Q201413","display_name":"Heuristic","level":2,"score":0.3379000127315521},{"id":"https://openalex.org/C183356978","wikidata":"https://www.wikidata.org/wiki/Q1779213","display_name":"Tracking error","level":3,"score":0.3375999927520752},{"id":"https://openalex.org/C55660270","wikidata":"https://www.wikidata.org/wiki/Q5164377","display_name":"Constrained optimization","level":2,"score":0.2924000024795532},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.2915000021457672},{"id":"https://openalex.org/C147764199","wikidata":"https://www.wikidata.org/wiki/Q6865248","display_name":"Minification","level":2,"score":0.2809999883174896},{"id":"https://openalex.org/C44616089","wikidata":"https://www.wikidata.org/wiki/Q30158686","display_name":"Constraint satisfaction","level":3,"score":0.2808000147342682},{"id":"https://openalex.org/C2776036281","wikidata":"https://www.wikidata.org/wiki/Q48769818","display_name":"Constraint (computer-aided design)","level":2,"score":0.2800000011920929},{"id":"https://openalex.org/C126388530","wikidata":"https://www.wikidata.org/wiki/Q1131737","display_name":"Imitation","level":2,"score":0.27889999747276306},{"id":"https://openalex.org/C181622380","wikidata":"https://www.wikidata.org/wiki/Q26911","display_name":"Profit (economics)","level":2,"score":0.26919999718666077},{"id":"https://openalex.org/C129844170","wikidata":"https://www.wikidata.org/wiki/Q41299","display_name":"Quadratic equation","level":2,"score":0.26570001244544983},{"id":"https://openalex.org/C95623464","wikidata":"https://www.wikidata.org/wiki/Q1096149","display_name":"Classifier (UML)","level":2,"score":0.26170000433921814},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.25940001010894775},{"id":"https://openalex.org/C77553402","wikidata":"https://www.wikidata.org/wiki/Q13222579","display_name":"Upper and lower bounds","level":2,"score":0.25600001215934753},{"id":"https://openalex.org/C137836250","wikidata":"https://www.wikidata.org/wiki/Q984063","display_name":"Optimization problem","level":2,"score":0.2549000084400177}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2604.14853","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.14853","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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.2604.14853","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.14853","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":"Quality Education","id":"https://metadata.un.org/sdg/4","score":0.49660369753837585}],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Test-time":[0],"compute":[1,71],"scaling,":[2],"the":[3,82,88,109,115,126,143,152,156,166,218],"practice":[4],"of":[5,155],"spending":[6],"extra":[7],"computation":[8],"during":[9],"inference":[10,35,176],"via":[11,122],"repeated":[12],"sampling,":[13],"search,":[14],"or":[15],"extended":[16],"reasoning,":[17],"has":[18],"become":[19],"a":[20,38,60,77,96,129,171],"powerful":[21],"lever":[22],"for":[23,146],"improving":[24],"large":[25],"language":[26],"model":[27],"performance.":[28],"Yet":[29],"deploying":[30],"these":[31],"techniques":[32],"under":[33,211],"finite":[34],"budgets":[36],"requires":[37],"decision":[39],"that":[40,100,108,151,192],"current":[41],"systems":[42],"largely":[43],"ignore:":[44],"which":[45,51],"inputs":[46],"deserve":[47],"more":[48],"compute,":[49],"and":[50,73,183,198],"can":[52],"be":[53],"answered":[54],"cheaply?":[55],"We":[56,106,149],"formalize":[57],"this":[58],"as":[59],"constrained":[61,175],"optimization":[62],"problem":[63],"(maximize":[64],"expected":[65],"accuracy":[66,103,207],"subject":[67],"to":[68,134,177,204],"an":[69],"average":[70],"budget)":[72],"solve":[74,83],"it":[75],"with":[76,185,223],"two-stage":[78],"Solve-then-Learn":[79],"pipeline.":[80],"In":[81,125],"stage,":[84,128],"Lagrangian":[85,219],"relaxation":[86],"decomposes":[87],"global":[89],"constraint":[90],"into":[91],"per-instance":[92,168],"sub-problems,":[93],"each":[94],"admitting":[95],"closed-form":[97],"oracle":[98,136,220],"action":[99],"optimally":[101],"prices":[102],"against":[104],"cost.":[105],"prove":[107],"induced":[110],"cost":[111],"is":[112,132,159],"monotone":[113],"in":[114],"dual":[116],"variable,":[117],"enabling":[118],"exact":[119],"budget":[120,213],"targeting":[121],"binary":[123],"search.":[124],"learn":[127],"lightweight":[130],"classifier":[131],"trained":[133],"predict":[135],"actions":[137],"from":[138,174],"cheap":[139],"input":[140],"features,":[141],"amortizing":[142],"allocation":[144,200],"rule":[145],"real-time":[147],"deployment.":[148],"establish":[150],"task-level":[153],"regret":[154],"learned":[157],"policy":[158],"bounded":[160],"by":[161],"its":[162],"imitation":[163,226],"error":[164],"times":[165],"worst-case":[167],"gap,":[169],"yielding":[170],"clean":[172],"reduction":[173],"supervised":[178],"classification.":[179],"Experiments":[180],"on":[181,209],"MATH":[182,210],"GSM8K":[184],"three":[186],"LLMs":[187],"(DeepSeek-V3,":[188],"GPT-4o-mini,":[189],"Qwen2.5-7B)":[190],"show":[191],"our":[193],"method":[194],"consistently":[195],"outperforms":[196],"uniform":[197],"heuristic":[199],"baselines,":[201],"achieving":[202],"up":[203],"12.8%":[205],"relative":[206],"improvement":[208],"matched":[212],"constraints,":[214],"while":[215],"closely":[216],"tracking":[217],"upper":[221],"bound":[222],"over":[224],"91%":[225],"accuracy.":[227]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-04-18T00:00:00"}