{"id":"https://openalex.org/W7164051778","doi":"https://doi.org/10.48550/arxiv.2606.08480","title":"Adaptive Loss Balancing for Noise-Robust GRPO in Generative Recommendation","display_name":"Adaptive Loss Balancing for Noise-Robust GRPO in Generative Recommendation","publication_year":2026,"publication_date":"2026-06-07","ids":{"openalex":"https://openalex.org/W7164051778","doi":"https://doi.org/10.48550/arxiv.2606.08480"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.08480","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.08480","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.2606.08480","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5122955745","display_name":"Kewei Xu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xu, Kewei","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138234923","display_name":"Junbo Qi","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Qi, Junbo","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138268456","display_name":"Yanyan Zou","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zou, Yanyan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138205714","display_name":"Pengfei Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Pengfei","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5127868513","display_name":"Xingzhi Yao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yao, Xingzhi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5138206198","display_name":"Shengjie Li (3340212)","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Shengjie","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/T10203","display_name":"Recommender Systems and Techniques","score":0.8384000062942505,"subfield":{"id":"https://openalex.org/subfields/1710","display_name":"Information Systems"},"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/T10203","display_name":"Recommender Systems and Techniques","score":0.8384000062942505,"subfield":{"id":"https://openalex.org/subfields/1710","display_name":"Information Systems"},"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":0.04390000179409981,"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/T12101","display_name":"Advanced Bandit Algorithms Research","score":0.00989999994635582,"subfield":{"id":"https://openalex.org/subfields/1803","display_name":"Management Science and Operations Research"},"field":{"id":"https://openalex.org/fields/18","display_name":"Decision Sciences"},"domain":{"id":"https://openalex.org/domains/2","display_name":"Social Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/robustness","display_name":"Robustness (evolution)","score":0.6862000226974487},{"id":"https://openalex.org/keywords/reinforcement-learning","display_name":"Reinforcement learning","score":0.6273000240325928},{"id":"https://openalex.org/keywords/trustworthiness","display_name":"Trustworthiness","score":0.5231000185012817},{"id":"https://openalex.org/keywords/generative-grammar","display_name":"Generative grammar","score":0.5169000029563904},{"id":"https://openalex.org/keywords/stability","display_name":"Stability (learning theory)","score":0.3734000027179718},{"id":"https://openalex.org/keywords/generative-model","display_name":"Generative model","score":0.3653999865055084}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6947000026702881},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.6862000226974487},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.6273000240325928},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.554099977016449},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.5396999716758728},{"id":"https://openalex.org/C153701036","wikidata":"https://www.wikidata.org/wiki/Q659974","display_name":"Trustworthiness","level":2,"score":0.5231000185012817},{"id":"https://openalex.org/C39890363","wikidata":"https://www.wikidata.org/wiki/Q36108","display_name":"Generative grammar","level":2,"score":0.5169000029563904},{"id":"https://openalex.org/C112972136","wikidata":"https://www.wikidata.org/wiki/Q7595718","display_name":"Stability (learning theory)","level":2,"score":0.3734000027179718},{"id":"https://openalex.org/C167966045","wikidata":"https://www.wikidata.org/wiki/Q5532625","display_name":"Generative model","level":3,"score":0.3653999865055084},{"id":"https://openalex.org/C136389625","wikidata":"https://www.wikidata.org/wiki/Q334384","display_name":"Supervised learning","level":3,"score":0.3483000099658966},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.3138999938964844},{"id":"https://openalex.org/C151637689","wikidata":"https://www.wikidata.org/wiki/Q5318064","display_name":"Dwell time","level":2,"score":0.3116999864578247},{"id":"https://openalex.org/C2778348673","wikidata":"https://www.wikidata.org/wiki/Q739302","display_name":"Production (economics)","level":2,"score":0.30489999055862427},{"id":"https://openalex.org/C125014702","wikidata":"https://www.wikidata.org/wiki/Q4680749","display_name":"Adaptive learning","level":2,"score":0.2897999882698059},{"id":"https://openalex.org/C67203356","wikidata":"https://www.wikidata.org/wiki/Q1321905","display_name":"Reinforcement","level":2,"score":0.2535000145435333}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.08480","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.08480","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.2606.08480","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.08480","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":"Reduced inequalities","id":"https://metadata.un.org/sdg/10","score":0.7055687308311462}],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Reinforcement":[0],"learning":[1],"(RL)":[2],"presents":[3],"a":[4,65,116,144,179],"promising":[5],"avenue":[6],"for":[7,34],"enhancing":[8],"generative":[9],"recommendation":[10],"beyond":[11],"supervised":[12,134],"imitation,":[13],"leveraging":[14],"reward":[15,32,46,68,94,156],"signals":[16],"to":[17,54,163,193],"guide":[18],"policy":[19,75],"improvement.":[20],"However,":[21],"its":[22,235],"efficacy":[23],"is":[24,70,96,131,141],"critically":[25],"contingent":[26],"on":[27,50,178],"the":[28,31,35,43,74,79,84,93,102,138,170,184,204,216],"trustworthiness":[29],"of":[30,104,172],"model":[33],"samples":[36],"it":[37,188],"evaluates.":[38],"In":[39,219],"practice,":[40],"production":[41,220],"rankers,":[42],"widely":[44],"adopted":[45],"models,":[47],"are":[48],"trained":[49],"exposure-biased":[51],"logs,":[52],"leading":[53],"sample-dependent":[55],"inaccuracies":[56],"that":[57,119],"violate":[58],"this":[59],"assumption.":[60],"Our":[61],"stratified":[62],"analysis":[63],"uncovers":[64],"consistent":[66],"pattern:":[67],"guidance":[69],"most":[71],"beneficial":[72],"when":[73],"exhibits":[76],"uncertainty":[77],"and":[78,155,168,200,231],"ranker":[80],"can":[81],"effectively":[82],"discriminate":[83],"ground-truth":[85],"item":[86],"from":[87,191],"rollout":[88,151],"negatives.":[89],"On":[90],"other":[91],"samples,":[92],"signal":[95],"either":[97,160],"negligible":[98],"or":[99],"detrimental,":[100],"highlighting":[101],"risk":[103],"uniform":[105,128],"RL":[106],"application.":[107],"To":[108],"address":[109],"such":[110],"an":[111],"issue,":[112],"we":[113],"introduce":[114],"AdaGRPO,":[115],"novel":[117],"framework":[118],"treats":[120],"reward-guided":[121],"optimization":[122],"as":[123],"selective":[124],"admission":[125],"rather":[126],"than":[127],"pressure.":[129],"Training":[130],"anchored":[132],"in":[133,228],"negative":[135],"log-likelihood,":[136],"while":[137,195],"GRPO":[139],"objective":[140],"gated":[142],"by":[143,149],"binary,":[145],"per-sample":[146],"clip":[147],"determined":[148],"two":[150],"diagnostics:":[152],"policy-side":[153],"difficulty":[154],"discriminability.":[157],"Instances":[158],"failing":[159],"diagnostic":[161],"default":[162],"pure":[164],"supervision,":[165],"ensuring":[166],"stability":[167],"mitigating":[169],"amplification":[171],"noisy":[173],"gradients.":[174],"We":[175],"validate":[176],"AdaGRPO":[177,223],"large-scale":[180],"e-commerce":[181],"dataset.":[182],"At":[183],"best":[185],"intermediate":[186],"checkpoint,":[187],"elevates":[189],"HR@10":[190],"11.01%":[192],"12.18%":[194],"constraining":[196],"hallucination":[197,209],"below":[198],"0.22%,":[199],"maintains":[201],"robustness":[202],"at":[203],"final":[205],"checkpoint":[206],"(HR@10":[207],"11.63%,":[208],"0.27%),":[210],"outperforming":[211],"fixed":[212],"NLL--GRPO":[213],"mixtures":[214],"across":[215],"retrieval--validity":[217],"frontier.":[218],"A/B":[221],"tests,":[222],"achieves":[224],"statistically":[225],"significant":[226],"gains":[227],"click-through":[229],"rate":[230],"dwell":[232],"time,":[233],"confirming":[234],"practical":[236],"utility.":[237]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-06-10T00:00:00"}