{"id":"https://openalex.org/W7164019967","doi":"https://doi.org/10.48550/arxiv.2606.08446","title":"Sparrow: Sparse Rollout for Stable and Efficient Long-context RL of Large Language Models","display_name":"Sparrow: Sparse Rollout for Stable and Efficient Long-context RL of Large Language Models","publication_year":2026,"publication_date":"2026-06-07","ids":{"openalex":"https://openalex.org/W7164019967","doi":"https://doi.org/10.48550/arxiv.2606.08446"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.08446","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.08446","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","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.2606.08446","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5138237231","display_name":"Yang Zhou","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhou, Yang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138219554","display_name":"Ranajoy Sadhukhan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Sadhukhan, Ranajoy","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138264728","display_name":"Zhaofeng Sun","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Sun, Zhaofeng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138226232","display_name":"Zhuoming Chen","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chen, Zhuoming","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138215401","display_name":"Souvik Kundu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Kundu, Souvik","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5027419911","display_name":"Saket Dingliwal","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Dingliwal, Saket","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5042855431","display_name":"Sai Muralidhar Jayanthi","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jayanthi, Sai Muralidhar","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101715504","display_name":"Aram Galstyan","orcid":"https://orcid.org/0000-0003-4215-0886"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Galstyan, Aram","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138204323","display_name":"Haizhong Zheng","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zheng, Haizhong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5138235076","display_name":"Beidi Chen","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chen, Beidi","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/T10028","display_name":"Topic Modeling","score":0.29170000553131104,"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":0.29170000553131104,"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/T11636","display_name":"Artificial Intelligence in Healthcare and Education","score":0.11969999969005585,"subfield":{"id":"https://openalex.org/subfields/2718","display_name":"Health Informatics"},"field":{"id":"https://openalex.org/fields/27","display_name":"Medicine"},"domain":{"id":"https://openalex.org/domains/4","display_name":"Health Sciences"}},{"id":"https://openalex.org/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.0860000029206276,"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/schedule","display_name":"Schedule","score":0.5407999753952026},{"id":"https://openalex.org/keywords/sparse-matrix","display_name":"Sparse matrix","score":0.47360000014305115},{"id":"https://openalex.org/keywords/sparse-approximation","display_name":"Sparse approximation","score":0.43149998784065247},{"id":"https://openalex.org/keywords/speedup","display_name":"Speedup","score":0.42179998755455017},{"id":"https://openalex.org/keywords/reinforcement-learning","display_name":"Reinforcement learning","score":0.39959999918937683},{"id":"https://openalex.org/keywords/domain","display_name":"Domain (mathematical analysis)","score":0.39320001006126404},{"id":"https://openalex.org/keywords/statistic","display_name":"Statistic","score":0.3813000023365021}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7383000254631042},{"id":"https://openalex.org/C68387754","wikidata":"https://www.wikidata.org/wiki/Q7271585","display_name":"Schedule","level":2,"score":0.5407999753952026},{"id":"https://openalex.org/C56372850","wikidata":"https://www.wikidata.org/wiki/Q1050404","display_name":"Sparse matrix","level":3,"score":0.47360000014305115},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.4607999920845032},{"id":"https://openalex.org/C124066611","wikidata":"https://www.wikidata.org/wiki/Q28684319","display_name":"Sparse approximation","level":2,"score":0.43149998784065247},{"id":"https://openalex.org/C68339613","wikidata":"https://www.wikidata.org/wiki/Q1549489","display_name":"Speedup","level":2,"score":0.42179998755455017},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.39959999918937683},{"id":"https://openalex.org/C36503486","wikidata":"https://www.wikidata.org/wiki/Q11235244","display_name":"Domain (mathematical analysis)","level":2,"score":0.39320001006126404},{"id":"https://openalex.org/C89128539","wikidata":"https://www.wikidata.org/wiki/Q1949963","display_name":"Statistic","level":2,"score":0.3813000023365021},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.33550000190734863},{"id":"https://openalex.org/C85847156","wikidata":"https://www.wikidata.org/wiki/Q59015987","display_name":"Verifiable secret sharing","level":3,"score":0.3319999873638153},{"id":"https://openalex.org/C2777027219","wikidata":"https://www.wikidata.org/wiki/Q1284190","display_name":"Constant (computer programming)","level":2,"score":0.2976999878883362},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.2953999936580658},{"id":"https://openalex.org/C147764199","wikidata":"https://www.wikidata.org/wiki/Q6865248","display_name":"Minification","level":2,"score":0.2815000116825104},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.28130000829696655},{"id":"https://openalex.org/C112972136","wikidata":"https://www.wikidata.org/wiki/Q7595718","display_name":"Stability (learning theory)","level":2,"score":0.28040000796318054},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.25940001010894775},{"id":"https://openalex.org/C2779960059","wikidata":"https://www.wikidata.org/wiki/Q7113681","display_name":"Overhead (engineering)","level":2,"score":0.2572000026702881},{"id":"https://openalex.org/C2779679103","wikidata":"https://www.wikidata.org/wiki/Q5251805","display_name":"Degradation (telecommunications)","level":2,"score":0.2549000084400177},{"id":"https://openalex.org/C5465570","wikidata":"https://www.wikidata.org/wiki/Q5326898","display_name":"Early stopping","level":3,"score":0.25049999356269836}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.08446","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.08446","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2606.08446","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.08446","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[{"display_name":"Decent work and economic growth","id":"https://metadata.un.org/sdg/8","score":0.45411449670791626}],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Despite":[0],"being":[1],"powerful,":[2],"reinforcement":[3],"learning":[4],"with":[5,88],"verifiable":[6],"rewards":[7],"(RLVR)":[8],"induces":[9],"extremely":[10],"long":[11],"COT,":[12],"making":[13],"it":[14],"computationally":[15],"expensive.":[16],"Since":[17],"RLVR":[18],"per-step":[19],"cost":[20,160],"is":[21,75],"dominated":[22],"by":[23,78,95],"long-context":[24],"rollout":[25,73,101,179,214],"generation,":[26],"sparse":[27,38,72,84,100,213],"attention":[28],"offers":[29],"a":[30,41,115,123,149,159,194],"promising":[31],"way":[32],"to":[33,162,193],"accelerate":[34],"dense":[35,89],"rollout.":[36],"However,":[37],"rollouts":[39],"require":[40],"delicate":[42],"stability-efficiency":[43],"tradeoff:":[44],"overly":[45,51],"aggressive":[46,92,217],"sparsity":[47,53,125,165,218],"causes":[48],"collapse,":[49],"while":[50],"lenient":[52],"gives":[54],"insufficient":[55],"speedup.":[56,227],"In":[57],"this":[58,62,129,171],"work,":[59],"we":[60,97,188],"study":[61],"tradeoff":[63],"through":[64],"sparse-to-dense":[65,222],"actor-policy":[66,111],"mismatch.":[67],"We":[68,121,156],"first":[69],"observe":[70],"that":[71,99,127],"collapse":[74],"not":[76],"driven":[77],"uniform":[79],"degradation":[80],"across":[81],"tokens:":[82],"most":[83],"tokens":[85],"align":[86],"perfectly":[87],"even":[90],"under":[91,170],"sparsity.":[93],"Motivated":[94],"this,":[96],"hypothesize":[98],"training":[102,182],"remains":[103],"stable":[104,154],"if":[105],"the":[106,119,144,164,190,220],"lower":[107],"tail":[108,130,145],"of":[109],"per-token":[110],"mismatch":[112,146,172,223],"stays":[113],"above":[114],"critical":[116],"threshold":[117,151],"throughout":[118],"trajectory.":[120],"introduce":[122],"dynamic":[124],"schedule":[126,166],"keeps":[128],"statistic":[131,147],"constant":[132],"during":[133],"generation":[134],"and":[135,177,185,198],"validate":[136],"our":[137,204],"hypothesis.":[138],"Across":[139],"Qwen3":[140],"thinking-family":[141],"models,":[142],"keeping":[143],"near":[148],"consistent":[150],"generally":[152],"enables":[153],"training.":[155],"then":[157],"use":[158],"model":[161,196],"find":[163],"for":[167],"maximum":[168],"speedup":[169],"threshold,":[173,224],"achieving":[174],"2.2x,":[175],"2.4x,":[176],"2.0x":[178],"speedups":[180],"when":[181],"Qwen3-1.7B,":[183],"Qwen3-4B,":[184],"Qwen3-8B.":[186],"Empirically,":[187],"show":[189],"thresholds":[191],"generalize":[192],"larger":[195],"(Qwen3-14B)":[197],"another":[199],"RL":[200],"domain":[201],"(coding).":[202],"Finally,":[203],"analysis":[205],"naturally":[206],"motivates":[207],"DistillSparse:":[208],"lightweight":[209],"LoRA-based":[210],"distillation":[211],"on":[212],"lets":[215],"more":[216],"reach":[219],"same":[221],"yielding":[225],"higher":[226]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-06-10T00:00:00"}