{"id":"https://openalex.org/W7160853126","doi":"https://doi.org/10.48550/arxiv.2605.06850","title":"How to Compress KV Cache in RL Post-Training? Shadow Mask Distillation for Memory-Efficient Alignment","display_name":"How to Compress KV Cache in RL Post-Training? Shadow Mask Distillation for Memory-Efficient Alignment","publication_year":2026,"publication_date":"2026-05-07","ids":{"openalex":"https://openalex.org/W7160853126","doi":"https://doi.org/10.48550/arxiv.2605.06850"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.06850","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.06850","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":"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.2605.06850","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5135900135","display_name":"Rui Zhu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhu, Rui","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5135887395","display_name":"Weiheng Bai","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Bai, Weiheng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5135873304","display_name":"Qiushi Wu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wu, Qiushi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5087938537","display_name":"Yang Ren","orcid":"https://orcid.org/0000-0002-6128-5826"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ren, Yang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5135859553","display_name":"Haixu Tang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tang, Haixu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5135886726","display_name":"Yuchu Liu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Yuchu","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":0.2079000025987625,"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":0.2079000025987625,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.12300000339746475,"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":0.08940000087022781,"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/lossless-compression","display_name":"Lossless compression","score":0.7250000238418579},{"id":"https://openalex.org/keywords/cache","display_name":"Cache","score":0.699400007724762},{"id":"https://openalex.org/keywords/sample","display_name":"Sample (material)","score":0.4925999939441681},{"id":"https://openalex.org/keywords/trajectory","display_name":"Trajectory","score":0.45419999957084656},{"id":"https://openalex.org/keywords/shadow","display_name":"Shadow (psychology)","score":0.4526999890804291},{"id":"https://openalex.org/keywords/reinforcement-learning","display_name":"Reinforcement learning","score":0.4440999925136566},{"id":"https://openalex.org/keywords/compression","display_name":"Compression (physics)","score":0.44209998846054077},{"id":"https://openalex.org/keywords/variance","display_name":"Variance (accounting)","score":0.40959998965263367}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.786899983882904},{"id":"https://openalex.org/C81081738","wikidata":"https://www.wikidata.org/wiki/Q55542","display_name":"Lossless compression","level":3,"score":0.7250000238418579},{"id":"https://openalex.org/C115537543","wikidata":"https://www.wikidata.org/wiki/Q165596","display_name":"Cache","level":2,"score":0.699400007724762},{"id":"https://openalex.org/C198531522","wikidata":"https://www.wikidata.org/wiki/Q485146","display_name":"Sample (material)","level":2,"score":0.4925999939441681},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.4702000021934509},{"id":"https://openalex.org/C13662910","wikidata":"https://www.wikidata.org/wiki/Q193139","display_name":"Trajectory","level":2,"score":0.45419999957084656},{"id":"https://openalex.org/C117797892","wikidata":"https://www.wikidata.org/wiki/Q286363","display_name":"Shadow (psychology)","level":2,"score":0.4526999890804291},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.4440999925136566},{"id":"https://openalex.org/C180016635","wikidata":"https://www.wikidata.org/wiki/Q2712821","display_name":"Compression (physics)","level":2,"score":0.44209998846054077},{"id":"https://openalex.org/C196083921","wikidata":"https://www.wikidata.org/wiki/Q7915758","display_name":"Variance (accounting)","level":2,"score":0.40959998965263367},{"id":"https://openalex.org/C189783530","wikidata":"https://www.wikidata.org/wiki/Q352090","display_name":"CPU cache","level":3,"score":0.3587000072002411},{"id":"https://openalex.org/C78548338","wikidata":"https://www.wikidata.org/wiki/Q2493","display_name":"Data compression","level":2,"score":0.3573000133037567},{"id":"https://openalex.org/C113775141","wikidata":"https://www.wikidata.org/wiki/Q428691","display_name":"Computer engineering","level":1,"score":0.3337000012397766},{"id":"https://openalex.org/C44280652","wikidata":"https://www.wikidata.org/wiki/Q104837","display_name":"Phase (matter)","level":2,"score":0.32109999656677246},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.29660001397132874},{"id":"https://openalex.org/C13481523","wikidata":"https://www.wikidata.org/wiki/Q412438","display_name":"Image compression","level":4,"score":0.29490000009536743},{"id":"https://openalex.org/C774472","wikidata":"https://www.wikidata.org/wiki/Q6760393","display_name":"Margin (machine learning)","level":2,"score":0.29190000891685486},{"id":"https://openalex.org/C112972136","wikidata":"https://www.wikidata.org/wiki/Q7595718","display_name":"Stability (learning theory)","level":2,"score":0.2897999882698059},{"id":"https://openalex.org/C127705205","wikidata":"https://www.wikidata.org/wiki/Q5748245","display_name":"Heuristics","level":2,"score":0.288100004196167},{"id":"https://openalex.org/C157764524","wikidata":"https://www.wikidata.org/wiki/Q1383412","display_name":"Throughput","level":3,"score":0.2782999873161316},{"id":"https://openalex.org/C165064840","wikidata":"https://www.wikidata.org/wiki/Q1321061","display_name":"Matching (statistics)","level":2,"score":0.2773999869823456},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.27059999108314514},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.262800008058548},{"id":"https://openalex.org/C2982736386","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Statistical learning","level":2,"score":0.2605000138282776}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.06850","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.06850","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":"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.2605.06850","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.06850","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":"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":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Reinforcement":[0],"Learning":[1],"(RL)":[2],"has":[3],"emerged":[4],"as":[5,134],"a":[6,57,82,117],"crucial":[7],"paradigm":[8],"for":[9,49],"unlocking":[10],"the":[11,28,63,105,112,121,126],"advanced":[12],"reasoning":[13,51],"capabilities":[14],"of":[15,27,108],"Large":[16],"Language":[17],"Models":[18],"(LLMs),":[19],"encompassing":[20],"frameworks":[21],"like":[22],"RLHF":[23],"and":[24,148],"RLAIF.":[25],"Regardless":[26],"specific":[29],"optimization":[30],"algorithm":[31],"(e.g.,":[32],"PPO,":[33],"GRPO,":[34],"or":[35],"Online":[36],"DPO),":[37],"online":[38],"RL":[39,109],"inherently":[40],"requires":[41],"an":[42],"exploratory":[43],"trajectory":[44],"generation":[45],"(rollout)":[46],"phase.":[47],"However,":[48],"long-context":[50],"tasks,":[52],"this":[53,77,140],"rollout":[54],"phase":[55],"imposes":[56],"severe":[58,149],"``memory":[59],"wall''":[60],"due":[61],"to":[62,138],"exorbitant":[64],"Key-Value":[65],"(KV)":[66],"cache":[67,72],"footprint.":[68],"While":[69],"applying":[70],"KV":[71,88],"compression":[73,89],"during":[74,94],"rollouts":[75],"mitigates":[76],"memory":[78],"overhead,":[79],"it":[80],"induces":[81],"critical":[83],"off-policy":[84],"bias.":[85],"Although":[86],"modern":[87],"is":[90],"often":[91],"nearly":[92],"lossless":[93],"standard":[95],"inference,":[96],"even":[97],"minuscule":[98],"approximation":[99],"errors":[100],"are":[101],"drastically":[102],"amplified":[103],"by":[104],"inherent":[106],"instability":[107],"optimization.":[110],"Specifically,":[111],"sampler":[113],"generates":[114],"responses":[115],"under":[116],"sparse":[118],"context,":[119],"whereas":[120],"learner":[122],"updates":[123],"parameters":[124],"using":[125],"full,":[127],"dense":[128],"context.":[129],"Existing":[130],"statistical":[131],"solutions,":[132],"such":[133],"importance":[135],"reweighting,":[136],"struggle":[137],"correct":[139],"magnified":[141],"bias,":[142],"suffering":[143],"from":[144],"high":[145],"gradient":[146],"variance":[147],"sample":[150],"inefficiency.":[151]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-05-12T00:00:00"}