{"id":"https://openalex.org/W7163019616","doi":"https://doi.org/10.48550/arxiv.2605.31455","title":"DRIFT: Decoupled Rollouts and Importance-Weighted Fine-Tuning for Efficient Multi-Turn Optimization","display_name":"DRIFT: Decoupled Rollouts and Importance-Weighted Fine-Tuning for Efficient Multi-Turn Optimization","publication_year":2026,"publication_date":"2026-05-29","ids":{"openalex":"https://openalex.org/W7163019616","doi":"https://doi.org/10.48550/arxiv.2605.31455"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.31455","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.31455","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":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.2605.31455","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5137520268","display_name":"Jian Mu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Mu, Jian","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137528036","display_name":"Tianyi Lin","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lin, Tianyi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137538003","display_name":"Chengwei Qin","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Qin, Chengwei","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137549392","display_name":"Zhongxiang Dai","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Dai, Zhongxiang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5137593648","display_name":"Yao Shu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Shu, Yao","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":0.4643999934196472,"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":0.4643999934196472,"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":0.09160000085830688,"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"}},{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.08560000360012054,"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/reinforcement-learning","display_name":"Reinforcement learning","score":0.7687000036239624},{"id":"https://openalex.org/keywords/code","display_name":"Code (set theory)","score":0.5562999844551086},{"id":"https://openalex.org/keywords/stability","display_name":"Stability (learning theory)","score":0.3589000105857849},{"id":"https://openalex.org/keywords/sampling","display_name":"Sampling (signal processing)","score":0.3587999939918518},{"id":"https://openalex.org/keywords/dilemma","display_name":"Dilemma","score":0.3587000072002411},{"id":"https://openalex.org/keywords/source-code","display_name":"Source code","score":0.35760000348091125},{"id":"https://openalex.org/keywords/concept-drift","display_name":"Concept drift","score":0.3449999988079071},{"id":"https://openalex.org/keywords/offline-learning","display_name":"Offline learning","score":0.3377000093460083}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.8009999990463257},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.7687000036239624},{"id":"https://openalex.org/C2776760102","wikidata":"https://www.wikidata.org/wiki/Q5139990","display_name":"Code (set theory)","level":3,"score":0.5562999844551086},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3919999897480011},{"id":"https://openalex.org/C112972136","wikidata":"https://www.wikidata.org/wiki/Q7595718","display_name":"Stability (learning theory)","level":2,"score":0.3589000105857849},{"id":"https://openalex.org/C140779682","wikidata":"https://www.wikidata.org/wiki/Q210868","display_name":"Sampling (signal processing)","level":3,"score":0.3587999939918518},{"id":"https://openalex.org/C2778496695","wikidata":"https://www.wikidata.org/wiki/Q254128","display_name":"Dilemma","level":2,"score":0.3587000072002411},{"id":"https://openalex.org/C43126263","wikidata":"https://www.wikidata.org/wiki/Q128751","display_name":"Source code","level":2,"score":0.35760000348091125},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3562000095844269},{"id":"https://openalex.org/C60777511","wikidata":"https://www.wikidata.org/wiki/Q3045002","display_name":"Concept drift","level":3,"score":0.3449999988079071},{"id":"https://openalex.org/C2780490138","wikidata":"https://www.wikidata.org/wiki/Q7079636","display_name":"Offline learning","level":3,"score":0.3377000093460083},{"id":"https://openalex.org/C2776372474","wikidata":"https://www.wikidata.org/wiki/Q508291","display_name":"Simplicity","level":2,"score":0.3314000070095062},{"id":"https://openalex.org/C2986087404","wikidata":"https://www.wikidata.org/wiki/Q15946010","display_name":"Online learning","level":2,"score":0.3253999948501587},{"id":"https://openalex.org/C137836250","wikidata":"https://www.wikidata.org/wiki/Q984063","display_name":"Optimization problem","level":2,"score":0.3222000002861023},{"id":"https://openalex.org/C136389625","wikidata":"https://www.wikidata.org/wiki/Q334384","display_name":"Supervised learning","level":3,"score":0.3197999894618988},{"id":"https://openalex.org/C52740198","wikidata":"https://www.wikidata.org/wiki/Q1539564","display_name":"Importance sampling","level":3,"score":0.30979999899864197},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.3093999922275543},{"id":"https://openalex.org/C13662910","wikidata":"https://www.wikidata.org/wiki/Q193139","display_name":"Trajectory","level":2,"score":0.2955999970436096},{"id":"https://openalex.org/C2777212361","wikidata":"https://www.wikidata.org/wiki/Q5127848","display_name":"Class (philosophy)","level":2,"score":0.28700000047683716},{"id":"https://openalex.org/C196921405","wikidata":"https://www.wikidata.org/wiki/Q786431","display_name":"Online algorithm","level":2,"score":0.28540000319480896},{"id":"https://openalex.org/C77618280","wikidata":"https://www.wikidata.org/wiki/Q1155772","display_name":"Scheme (mathematics)","level":2,"score":0.27219998836517334},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.25949999690055847},{"id":"https://openalex.org/C110121322","wikidata":"https://www.wikidata.org/wiki/Q865811","display_name":"Distribution (mathematics)","level":2,"score":0.2535000145435333}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.31455","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.31455","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":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.2605.31455","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.31455","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Large":[0],"language":[1],"models":[2],"are":[3],"increasingly":[4],"deployed":[5],"in":[6,27],"multi-turn":[7,37,141],"interactive":[8],"settings":[9],"where":[10],"users":[11],"or":[12,136],"environments":[13],"can":[14],"iteratively":[15],"provide":[16],"lightweight":[17],"feedback.":[18],"Unfortunately,":[19],"optimizing":[20,120],"such":[21],"behavior":[22],"presents":[23],"a":[24,82,111],"sharp":[25],"dilemma":[26],"practice:":[28],"online":[29],"reinforcement":[30,142],"learning":[31,143],"is":[32,40,60,94,157],"able":[33],"to":[34,44,96],"effectively":[35],"address":[36],"dynamics":[38],"but":[39,62],"prohibitively":[41],"expensive":[42],"due":[43],"the":[45,86,90,121,127,138,147],"cost":[46],"of":[47,140,152],"generating":[48],"full":[49],"correction":[50],"trajectories":[51,109],"at":[52,159],"every":[53],"update,":[54],"whereas":[55],"offline":[56,107],"supervised":[57,98,154],"fine-tuning":[58],"(SFT)":[59],"efficient":[61],"suffers":[63],"from":[64,103,110],"distribution":[65],"shift":[66],"and":[67,79,119,150],"behavioral":[68],"collapse.":[69],"To":[70],"this":[71],"end,":[72],"we":[73,131],"novelly":[74],"propose":[75],"DRIFT":[76,100,134],"(Decoupled":[77],"Rollouts":[78],"Importance-Weighted":[80],"Fine-Tuning),":[81],"framework":[83],"that":[84,89,133],"operationalizes":[85],"theoretical":[87],"insight":[88],"KL-regularized":[91],"RL":[92],"objective":[93],"equivalent":[95],"importance-weighted":[97],"learning.":[99],"decouples":[101],"rollout":[102],"optimization":[104],"by":[105],"sampling":[106],"interaction":[108],"fixed":[112],"reference":[113],"policy,":[114],"deriving":[115],"return-based":[116],"importance":[117],"weights,":[118],"policy":[122],"via":[123],"weighted":[124],"SFT":[125],"on":[126],"resulting":[128],"dataset.":[129],"Empirically,":[130],"demonstrate":[132],"matches":[135],"exceeds":[137],"performance":[139],"baselines":[144],"while":[145],"maintaining":[146],"training":[148],"efficiency":[149],"simplicity":[151],"standard":[153],"fine-tuning.":[155],"Code":[156],"available":[158],"https://github.com/2020-qqtcg/DRIFT.":[160]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-06-02T00:00:00"}