{"id":"https://openalex.org/W7140126924","doi":"https://doi.org/10.48550/arxiv.2603.19266","title":"Probing to Refine: Reinforcement Distillation of LLMs via Explanatory Inversion","display_name":"Probing to Refine: Reinforcement Distillation of LLMs via Explanatory Inversion","publication_year":2026,"publication_date":"2026-02-26","ids":{"openalex":"https://openalex.org/W7140126924","doi":"https://doi.org/10.48550/arxiv.2603.19266"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2603.19266","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.19266","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.2603.19266","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5130396107","display_name":"Zhen Tan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tan, Zhen","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5130365600","display_name":"Chengshuai Zhao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhao, Chengshuai","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5130363804","display_name":"Song Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Song","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5130351804","display_name":"Jundong Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Jundong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5130360247","display_name":"Tianlong Chen","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chen, Tianlong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5130347078","display_name":"Huan Liu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Huan","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":0.5695000290870667,"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":0.5695000290870667,"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":0.1517000049352646,"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":0.07429999858140945,"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/distillation","display_name":"Distillation","score":0.7692999839782715},{"id":"https://openalex.org/keywords/inversion","display_name":"Inversion (geology)","score":0.5806000232696533},{"id":"https://openalex.org/keywords/generalization","display_name":"Generalization","score":0.5526000261306763},{"id":"https://openalex.org/keywords/memorization","display_name":"Memorization","score":0.5203999876976013},{"id":"https://openalex.org/keywords/reinforcement-learning","display_name":"Reinforcement learning","score":0.5037999749183655},{"id":"https://openalex.org/keywords/process","display_name":"Process (computing)","score":0.4799000024795532},{"id":"https://openalex.org/keywords/simple","display_name":"Simple (philosophy)","score":0.4277999997138977}],"concepts":[{"id":"https://openalex.org/C204030448","wikidata":"https://www.wikidata.org/wiki/Q101017","display_name":"Distillation","level":2,"score":0.7692999839782715},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.6155999898910522},{"id":"https://openalex.org/C1893757","wikidata":"https://www.wikidata.org/wiki/Q3653001","display_name":"Inversion (geology)","level":3,"score":0.5806000232696533},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5742999911308289},{"id":"https://openalex.org/C177148314","wikidata":"https://www.wikidata.org/wiki/Q170084","display_name":"Generalization","level":2,"score":0.5526000261306763},{"id":"https://openalex.org/C30038468","wikidata":"https://www.wikidata.org/wiki/Q4354775","display_name":"Memorization","level":2,"score":0.5203999876976013},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.5123999714851379},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.5037999749183655},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.4799000024795532},{"id":"https://openalex.org/C2780586882","wikidata":"https://www.wikidata.org/wiki/Q7520643","display_name":"Simple (philosophy)","level":2,"score":0.4277999997138977},{"id":"https://openalex.org/C7863114","wikidata":"https://www.wikidata.org/wiki/Q192627","display_name":"Mimicry","level":2,"score":0.3206000030040741},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.32019999623298645},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.29269999265670776},{"id":"https://openalex.org/C117765406","wikidata":"https://www.wikidata.org/wiki/Q5362437","display_name":"Generalization error","level":3,"score":0.2924000024795532},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.290800005197525},{"id":"https://openalex.org/C67203356","wikidata":"https://www.wikidata.org/wiki/Q1321905","display_name":"Reinforcement","level":2,"score":0.2831000089645386},{"id":"https://openalex.org/C177606310","wikidata":"https://www.wikidata.org/wiki/Q5674297","display_name":"Adaptability","level":2,"score":0.2612999975681305},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.25850000977516174},{"id":"https://openalex.org/C196083921","wikidata":"https://www.wikidata.org/wiki/Q7915758","display_name":"Variance (accounting)","level":2,"score":0.2565999925136566}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2603.19266","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.19266","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.2603.19266","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.19266","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":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Distilling":[0],"robust":[1],"reasoning":[2,117],"capabilities":[3],"from":[4,26],"large":[5],"language":[6],"models":[7,14,23,156],"(LLMs)":[8],"into":[9],"smaller,":[10],"computationally":[11],"efficient":[12],"student":[13,75,112,134],"remains":[15],"an":[16,82,139],"unresolved":[17],"challenge.":[18],"Despite":[19],"recent":[20],"advances,":[21],"distilled":[22,157],"frequently":[24],"suffer":[25],"superficial":[27],"pattern":[28,63],"memorization":[29],"and":[30,146,173],"subpar":[31],"generalization.":[32],"To":[33],"overcome":[34],"these":[35,120],"limitations,":[36],"we":[37],"introduce":[38],"a":[39,50,97,102,115,147],"novel":[40,103],"distillation":[41,153],"framework":[42,55],"that":[43,72],"moves":[44],"beyond":[45],"simple":[46],"mimicry":[47],"to":[48,61,76,90,176],"instill":[49],"deeper":[51],"conceptual":[52],"understanding.":[53],"Our":[54],"features":[56],"two":[57],"key":[58],"innovations.":[59],"\\underline{\\textit{First}},":[60],"address":[62],"memorization,":[64],"Explanatory":[65,93],"Inversion":[66],"(EI)":[67],"generates":[68],"targeted":[69],"``explanatory":[70],"probes''":[71],"compel":[73],"the":[74,78,111,133,151],"articulate":[77],"underlying":[79],"logic":[80],"behind":[81],"answer,":[83],"rather":[84],"than":[85],"just":[86],"memorizing":[87],"it.":[88],"\\underline{\\textit{Second}},":[89],"improve":[91],"generalization,":[92],"GRPO":[94],"(\\texttt{EXGRPO})":[95],"uses":[96],"reinforcement":[98],"learning":[99],"algorithm":[100],"with":[101,158,169],"Dialogue":[104],"Structure":[105],"Utility":[106],"Bonus,":[107],"which":[108],"explicitly":[109],"rewards":[110],"for":[113],"maintaining":[114],"coherent":[116],"process":[118],"across":[119],"probes.":[121],"Extensive":[122],"evaluations":[123],"on":[124],"12":[125],"datasets":[126],"demonstrate":[127],"significant":[128],"improvements.":[129],"Using":[130],"Gemma-7b":[131],"as":[132],"model,":[135],"our":[136,159],"method":[137,160],"yields":[138],"average":[140],"\\textbf{20.39\\%}":[141],"increase":[142],"over":[143,150],"zero-shot":[144],"performance":[145],"\\textbf{6.02\\%}":[148],"improvement":[149],"state-of-the-art":[152],"baselines.":[154],"Moreover,":[155],"show":[161],"remarkable":[162],"training":[163,171],"efficiency":[164],"(e.g.,":[165],"surpassing":[166],"vanilla":[167],"fine-tuning":[168],"\\textbf{10-25\\%}":[170],"data)":[172],"strong":[174],"generalization":[175],"out-of-distribution":[177],"tasks.":[178],"Implementation":[179],"is":[180],"released":[181],"at":[182],"https://github.com/Zhen-Tan-dmml/ExGRPO.git.":[183]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-03-24T00:00:00"}