{"id":"https://openalex.org/W7214769125","doi":"https://doi.org/10.48550/arxiv.2609.35210","title":"Understanding On-Policy Distillation: A Mechanistic Interpretability Perspective via Sparse Crosscoders","display_name":"Understanding On-Policy Distillation: A Mechanistic Interpretability Perspective via Sparse Crosscoders","publication_year":2026,"publication_date":"2026-09-28","ids":{"openalex":"https://openalex.org/W7214769125","doi":"https://doi.org/10.48550/arxiv.2609.35210"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2609.35210","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.35210","pdf_url":"https://arxiv.org/pdf/2609.35210","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,"listed_in":[],"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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2609.35210","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5112840113","display_name":"Zichao Yu","orcid":"https://orcid.org/0009-0005-8880-279X","observed_orcids":["https://orcid.org/0009-0005-8880-279X"]},"institutions":[{"id":"https://openalex.org/I889458895","display_name":"University of Hong Kong","ror":"https://ror.org/02zhqgq86","country_code":"HK","type":"education","lineage":["https://openalex.org/I889458895"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Yu, Zichao","raw_affiliation_strings":["School of Computing and Data Science , The University of Hong Kong"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Computing and Data Science , The University of Hong Kong","institution_ids":["https://openalex.org/I889458895"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5005658216","display_name":"Qilang Ye","orcid":"https://orcid.org/0009-0004-6907-702X","observed_orcids":["https://orcid.org/0009-0004-6907-702X"]},"institutions":[{"id":"https://openalex.org/I241749","display_name":"University of Cambridge","ror":"https://ror.org/013meh722","country_code":"GB","type":"education","lineage":["https://openalex.org/I241749"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Ye, Qianshuo","raw_affiliation_strings":["Department of Computer Science and Technology , University of Cambridge"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Computer Science and Technology , University of Cambridge","institution_ids":["https://openalex.org/I241749"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5153531033","display_name":"Xu Wang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I889458895","display_name":"University of Hong Kong","ror":"https://ror.org/02zhqgq86","country_code":"HK","type":"education","lineage":["https://openalex.org/I889458895"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Wang, Xu","raw_affiliation_strings":["School of Computing and Data Science , The University of Hong Kong","Shenzhen Loop Area Institute"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Computing and Data Science , The University of Hong Kong","institution_ids":["https://openalex.org/I889458895"]},{"raw_affiliation_string":"Shenzhen Loop Area Institute","institution_ids":[]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5085848346","display_name":"Difan Zou","orcid":"https://orcid.org/0000-0002-6544-2593","observed_orcids":["https://orcid.org/0000-0002-6544-2593"]},"institutions":[{"id":"https://openalex.org/I889458895","display_name":"University of Hong Kong","ror":"https://ror.org/02zhqgq86","country_code":"HK","type":"education","lineage":["https://openalex.org/I889458895"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Zou, Difan","raw_affiliation_strings":["School of Computing and Data Science , The University of Hong Kong","Shenzhen Loop Area Institute"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Computing and Data Science , The University of Hong Kong","institution_ids":["https://openalex.org/I889458895"]},{"raw_affiliation_string":"Shenzhen Loop Area Institute","institution_ids":[]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":2,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":1,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.9818999767303467,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.9818999767303467,"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/T10028","display_name":"Topic Modeling","score":0.007000000216066837,"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":0.004900000058114529,"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/on-policy-distillation","display_name":"on-policy distillation","score":1.0},{"id":"https://openalex.org/keywords/feature-dictionary","display_name":"feature dictionary","score":0.4090000092983246},{"id":"https://openalex.org/keywords/feature-reweighting","display_name":"feature reweighting","score":0.6620000004768372},{"id":"https://openalex.org/keywords/teacher-student-knowledge-transfer","display_name":"teacher-student knowledge transfer","score":0.7419999837875366},{"id":"https://openalex.org/keywords/llm-reasoning","display_name":"LLM reasoning","score":0.6150000095367432},{"id":"https://openalex.org/keywords/mechanistic-interpretability","display_name":"mechanistic interpretability","score":0.5709999799728394}],"concepts":[{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.8147000074386597},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6477000117301941},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.6251000165939331},{"id":"https://openalex.org/C12713177","wikidata":"https://www.wikidata.org/wiki/Q1900281","display_name":"Perspective (graphical)","level":2,"score":0.6061999797821045},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5899999737739563},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.5224999785423279},{"id":"https://openalex.org/C99821215","wikidata":"https://www.wikidata.org/wiki/Q1136583","display_name":"Swap (finance)","level":2,"score":0.48809999227523804},{"id":"https://openalex.org/C127705205","wikidata":"https://www.wikidata.org/wiki/Q5748245","display_name":"Heuristics","level":2,"score":0.45579999685287476},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.4052000045776367},{"id":"https://openalex.org/C2777200299","wikidata":"https://www.wikidata.org/wiki/Q52943","display_name":"Conversation","level":2,"score":0.36480000615119934},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.36399999260902405},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.3246000111103058},{"id":"https://openalex.org/C77637269","wikidata":"https://www.wikidata.org/wiki/Q7002051","display_name":"Neural coding","level":2,"score":0.3111000061035156},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.2612000107765198},{"id":"https://openalex.org/C150899416","wikidata":"https://www.wikidata.org/wiki/Q1820378","display_name":"Transfer of learning","level":2,"score":0.25429999828338623}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2609.35210","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.35210","pdf_url":"https://arxiv.org/pdf/2609.35210","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,"listed_in":[],"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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2609.35210","doi":"https://doi.org/10.48550/arxiv.2609.35210","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.35210","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,"listed_in":[],"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"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2609.35210","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.35210","pdf_url":"https://arxiv.org/pdf/2609.35210","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,"listed_in":[],"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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[{"display_name":"Quality education","id":"https://metadata.un.org/sdg/4","score":0.5385000109672546}],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7214769125.pdf","grobid_xml":"https://content.openalex.org/works/W7214769125.grobid-xml"},"referenced_works_count":22,"referenced_works":["https://openalex.org/W3168867926","https://openalex.org/W4403883066","https://openalex.org/W4405254162","https://openalex.org/W4414281281","https://openalex.org/W4414939355","https://openalex.org/W4415566463","https://openalex.org/W4416033956","https://openalex.org/W4416131312","https://openalex.org/W4416600223","https://openalex.org/W4417530951","https://openalex.org/W7118867773","https://openalex.org/W7130402983","https://openalex.org/W7153282135","https://openalex.org/W7154505894","https://openalex.org/W7161054869","https://openalex.org/W7162992044","https://openalex.org/W7164026309","https://openalex.org/W7171552453","https://openalex.org/W7202118241","https://openalex.org/W7202205122","https://openalex.org/W7202369261","https://openalex.org/W7204908525"],"related_works":["https://openalex.org/W7214987755","https://openalex.org/W7154505894","https://openalex.org/W7162322277","https://openalex.org/W7220404346","https://openalex.org/W7169504124","https://openalex.org/W7163185507","https://openalex.org/W7171778044","https://openalex.org/W7168964786","https://openalex.org/W7164013384","https://openalex.org/W7161054869"],"abstract_inverted_index":{"On-policy":[0],"distillation":[1],"(OPD)":[2],"is":[3,13],"a":[4,20,69,237],"widely":[5],"adopted":[6],"post-training":[7],"technique":[8],"for":[9,114,219],"LLM":[10],"reasoning.":[11],"It":[12],"commonly":[14,164],"believed":[15],"to":[16,250,286],"transfer":[17],"knowledge":[18],"from":[19,282],"stronger":[21],"teacher,":[22],"yet":[23],"what":[24],"OPD":[25,54,122,127,166,196,213,270],"actually":[26],"distills":[27],"into":[28,81],"the":[29,49,56,90,107,118,134,139,146,156,160,176,179,193,253,257,279,283,288],"student's":[30,108,147,240],"internal":[31],"representations":[32],"remains":[33],"unclear.":[34],"We":[35,87,153],"study":[36],"this":[37,234],"question":[38],"with":[39],"sparse":[40],"crosscoders,":[41],"which":[42,93],"learn":[43],"one":[44,82],"feature":[45,85,98],"dictionary":[46],"shared":[47,180],"by":[48,117],"student":[50,96,280],"before":[51],"and":[52,55,137,167,189,199,224,227],"after":[53],"teacher.":[57],"Standard":[58],"crosscoder":[59],"analyses,":[60],"however,":[61],"identify":[62],"model-specific":[63],"features":[64,74,130,150,194,211,262,273,289],"but":[65],"cannot":[66],"tell":[67],"how":[68,104,285],"model's":[70],"use":[71,109,287],"of":[72,84,110,142,145,192,203,252],"its":[73,101,244,247],"changes,":[75],"since":[76],"all":[77],"models":[78],"are":[79],"encoded":[80],"set":[83],"activations.":[86],"therefore":[88],"propose":[89],"swap":[91],"readout,":[92],"reads":[94],"each":[95,111],"checkpoint's":[97],"activations":[99],"on":[100,133,159,236,260],"own,":[102,136],"measuring":[103],"training":[105],"changes":[106,210,229],"feature,":[112],"even":[113],"checkpoints":[115],"unseen":[116],"crosscoder.":[119],"Across":[120],"three":[121],"settings,":[123],"we":[124],"find":[125],"that":[126,163,195,212,251,269],"neither":[128],"creates":[129],"nor":[131],"passes":[132],"teacher's":[135,161],"leaves":[138],"firing":[140],"rates":[141],"over":[143],"98%":[144],"frequently":[148],"used":[149],"within":[151],"20%.":[152],"further":[154],"examine":[155],"SFT":[157],"warm-up":[158,177],"rollouts":[162],"precedes":[165],"makes":[168],"it":[169,186,209],"more":[170],"effective.":[171],"Rather":[172],"than":[173,275],"adding":[174],"features,":[175,241],"reweights":[178,271],"ones":[181],"in":[182,206],"two":[183],"ways.":[184],"First,":[185],"already":[187,291],"raises":[188,198],"lowers":[190],"many":[191],"later":[197],"lowers,":[200],"doing":[201],"part":[202],"OPD's":[204],"work":[205],"advance.":[207],"Second,":[208],"alone":[214],"would":[215],"not,":[216],"notably":[217],"those":[218],"conversation":[220],"format,":[221],"reasoning":[222],"style,":[223],"mathematical":[225],"notation,":[226],"these":[228,266],"persist":[230],"through":[231],"OPD.":[232],"Imposing":[233],"reweighting":[235],"directly":[238],"distilled":[239],"without":[242],"changing":[243],"weights,":[245],"brings":[246],"accuracy":[248],"close":[249],"warmed-up":[254],"student,":[255],"whereas":[256],"same":[258],"change":[259],"shuffled":[261],"does":[263],"not.":[264],"Together,":[265],"findings":[267],"suggest":[268],"existing":[272],"rather":[274],"acquiring":[276],"new":[277],"ones:":[278],"learns":[281],"teacher":[284],"they":[290],"share.":[292]},"counts_by_year":[{"year":2026,"cited_by_count":1}],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-09-30T00:00:00"}