{"id":"https://openalex.org/W7160876130","doi":"https://doi.org/10.48550/arxiv.2605.07660","title":"Not All Tokens Learn Alike: Attention Entropy Reveals Heterogeneous Signals in RL Reasoning","display_name":"Not All Tokens Learn Alike: Attention Entropy Reveals Heterogeneous Signals in RL Reasoning","publication_year":2026,"publication_date":"2026-05-08","ids":{"openalex":"https://openalex.org/W7160876130","doi":"https://doi.org/10.48550/arxiv.2605.07660"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.07660","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.07660","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.2605.07660","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5122323955","display_name":"Gengyang Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Gengyang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5135877794","display_name":"Zheng-Fan Wu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wu, Zheng-Fan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5135870357","display_name":"Siqi Bao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Bao, Siqi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5135853400","display_name":"Yunfang Wu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wu, Yunfang","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":0.30790001153945923,"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":0.30790001153945923,"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":0.2021999955177307,"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":0.12099999934434891,"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/security-token","display_name":"Security token","score":0.5824000239372253},{"id":"https://openalex.org/keywords/entropy","display_name":"Entropy (arrow of time)","score":0.57669997215271},{"id":"https://openalex.org/keywords/normalization","display_name":"Normalization (sociology)","score":0.4542999863624573},{"id":"https://openalex.org/keywords/kullback\u2013leibler-divergence","display_name":"Kullback\u2013Leibler divergence","score":0.38960000872612},{"id":"https://openalex.org/keywords/optimization-problem","display_name":"Optimization problem","score":0.3734999895095825},{"id":"https://openalex.org/keywords/bayesian-probability","display_name":"Bayesian probability","score":0.33889999985694885}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6420999765396118},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.5824000239372253},{"id":"https://openalex.org/C106301342","wikidata":"https://www.wikidata.org/wiki/Q4117933","display_name":"Entropy (arrow of time)","level":2,"score":0.57669997215271},{"id":"https://openalex.org/C136886441","wikidata":"https://www.wikidata.org/wiki/Q926129","display_name":"Normalization (sociology)","level":2,"score":0.4542999863624573},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.43970000743865967},{"id":"https://openalex.org/C171752962","wikidata":"https://www.wikidata.org/wiki/Q255166","display_name":"Kullback\u2013Leibler divergence","level":2,"score":0.38960000872612},{"id":"https://openalex.org/C137836250","wikidata":"https://www.wikidata.org/wiki/Q984063","display_name":"Optimization problem","level":2,"score":0.3734999895095825},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.36640000343322754},{"id":"https://openalex.org/C107673813","wikidata":"https://www.wikidata.org/wiki/Q812534","display_name":"Bayesian probability","level":2,"score":0.33889999985694885},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.3312999904155731},{"id":"https://openalex.org/C152124472","wikidata":"https://www.wikidata.org/wiki/Q1204361","display_name":"Redundancy (engineering)","level":2,"score":0.29409998655319214},{"id":"https://openalex.org/C117220453","wikidata":"https://www.wikidata.org/wiki/Q5172842","display_name":"Correlation","level":2,"score":0.28529998660087585},{"id":"https://openalex.org/C4679612","wikidata":"https://www.wikidata.org/wiki/Q866298","display_name":"Aggregate (composite)","level":2,"score":0.2797999978065491},{"id":"https://openalex.org/C2778049539","wikidata":"https://www.wikidata.org/wiki/Q17002908","display_name":"Bayesian optimization","level":2,"score":0.26829999685287476}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.07660","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.07660","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.2605.07660","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.07660","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":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Reinforcement-learning-based":[0],"post-training":[1],"has":[2],"become":[3],"a":[4,100,180],"key":[5],"approach":[6],"for":[7,42],"improving":[8],"the":[9,38,65,176,194],"reasoning":[10,220],"ability":[11],"of":[12,64],"large":[13],"language":[14],"models,":[15],"but":[16,104,124],"its":[17],"token-level":[18,50,73,207],"learning":[19],"signals":[20,146],"remain":[21],"poorly":[22],"understood.":[23],"This":[24],"work":[25],"studies":[26],"their":[27],"heterogeneity":[28,218],"through":[29],"attention":[30,201],"entropy,":[31,169],"which":[32,83,113],"measures":[33],"how":[34],"concentrated":[35,89],"or":[36],"diffuse":[37,119],"contextual":[39],"support":[40,152],"is":[41,130],"each":[43],"response":[44],"token.":[45],"We":[46,151],"first":[47],"show":[48],"that":[49,139,166,200,211],"RL":[51,208],"objectives":[52],"are":[53],"sparsely":[54],"estimable:":[55],"uniformly":[56],"random":[57],"20":[58],"percent":[59],"token":[60,213],"subsets":[61,77],"preserve":[62],"much":[63],"full-token":[66,96],"held-out":[67,191],"performance,":[68],"suggesting":[69],"substantial":[70],"redundancy":[71],"in":[72,193,206,219],"updates.":[74],"However,":[75],"entropy-structured":[76],"behave":[78],"very":[79],"differently.":[80],"Low-attention-entropy":[81],"tokens,":[82,112],"we":[84,114],"call":[85,115],"anchors,":[86],"rely":[87],"on":[88,108,132],"support,":[90],"produce":[91],"stable":[92],"gradients":[93],"aligned":[94],"with":[95,156],"updates,":[97],"and":[98,121,163,170,210],"provide":[99],"reliable":[101],"optimization":[102,148],"backbone,":[103],"tend":[105],"to":[106,189],"plateau":[107],"harder":[109],"benchmarks.":[110],"High-attention-entropy":[111],"explorers,":[116],"aggregate":[117],"more":[118,125],"context":[120],"induce":[122],"larger":[123],"volatile":[126],"gradients.":[127],"Explorer-only":[128],"training":[129],"unstable":[131],"average,":[133],"though":[134],"rare":[135],"successful":[136],"runs":[137],"suggest":[138,199],"these":[140],"tokens":[141],"may":[142],"contain":[143],"useful":[144],"hard-reasoning":[145],"when":[147],"remains":[149],"stable.":[150],"this":[153],"anchor-explorer":[154],"spectrum":[155],"evidence-gathering":[157],"analyses,":[158],"entropy":[159,202],"dynamics,":[160],"gradient-geometry":[161],"diagnostics,":[162],"controls":[164],"showing":[165],"position,":[167],"predictive":[168],"loss":[171],"normalization":[172],"do":[173],"not":[174],"explain":[175],"observed":[177],"asymmetry.":[178],"Finally,":[179],"dynamic":[181],"entropy-aware":[182],"soft-reweighting":[183],"intervention":[184],"improves":[185],"Qwen3-8B-Base":[186],"from":[187],"34.39":[188],"37.40":[190],"average":[192],"strongest":[195],"setting.":[196],"These":[197],"findings":[198],"reveals":[203],"optimization-relevant":[204],"structure":[205],"signals,":[209],"uniform":[212],"averaging":[214],"can":[215],"obscure":[216],"meaningful":[217],"post-training.":[221]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-05-12T00:00:00"}