{"id":"https://openalex.org/W7216460569","doi":"https://doi.org/10.48550/arxiv.2610.00694","title":"How Divergence Becomes Decision Flips in Compressed Language Models","display_name":"How Divergence Becomes Decision Flips in Compressed Language Models","publication_year":2026,"publication_date":"2026-09-30","ids":{"openalex":"https://openalex.org/W7216460569","doi":"https://doi.org/10.48550/arxiv.2610.00694"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2610.00694","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.00694","pdf_url":"https://arxiv.org/pdf/2610.00694","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/2610.00694","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5144248237","display_name":"Beatriz Almeida Fel\u00edcio","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Felicio, Beatriz Almeida","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":["https://openalex.org/A5144248237"],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"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":0.5692999958992004,"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":0.5692999958992004,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.1639000028371811,"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":0.10970000177621841,"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/language-model-compression","display_name":"language model compression","score":0.7860000133514404},{"id":"https://openalex.org/keywords/kullback-leibler-divergence","display_name":"Kullback-Leibler divergence","score":0.7089999914169312},{"id":"https://openalex.org/keywords/total-variation","display_name":"total variation","score":0.9380000233650208},{"id":"https://openalex.org/keywords/divergence-measures","display_name":"divergence measures","score":0.5170000195503235},{"id":"https://openalex.org/keywords/hellinger-distance","display_name":"Hellinger distance","score":0.8429999947547913},{"id":"https://openalex.org/keywords/perturbation-analysis","display_name":"perturbation analysis","score":0.43700000643730164},{"id":"https://openalex.org/keywords/teacher-forcing","display_name":"teacher forcing","score":0.7900000214576721},{"id":"https://openalex.org/keywords/code-generation","display_name":"code generation","score":0.42899999022483826}],"concepts":[{"id":"https://openalex.org/C207390915","wikidata":"https://www.wikidata.org/wiki/Q1230525","display_name":"Divergence (linguistics)","level":2,"score":0.5633999705314636},{"id":"https://openalex.org/C105795698","wikidata":"https://www.wikidata.org/wiki/Q12483","display_name":"Statistics","level":1,"score":0.5120000243186951},{"id":"https://openalex.org/C62100291","wikidata":"https://www.wikidata.org/wiki/Q1936288","display_name":"Total variation","level":2,"score":0.5099999904632568},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.46959999203681946},{"id":"https://openalex.org/C2778334786","wikidata":"https://www.wikidata.org/wiki/Q1586270","display_name":"Variation (astronomy)","level":2,"score":0.44510000944137573},{"id":"https://openalex.org/C180016635","wikidata":"https://www.wikidata.org/wiki/Q2712821","display_name":"Compression (physics)","level":2,"score":0.4390000104904175},{"id":"https://openalex.org/C153024298","wikidata":"https://www.wikidata.org/wiki/Q3030678","display_name":"Hellinger distance","level":2,"score":0.4153999984264374},{"id":"https://openalex.org/C139945424","wikidata":"https://www.wikidata.org/wiki/Q1940696","display_name":"Mean squared error","level":2,"score":0.38519999384880066},{"id":"https://openalex.org/C197115733","wikidata":"https://www.wikidata.org/wiki/Q1003136","display_name":"Forcing (mathematics)","level":2,"score":0.37450000643730164},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.3698999881744385},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.36629998683929443},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.36500000953674316},{"id":"https://openalex.org/C2776760102","wikidata":"https://www.wikidata.org/wiki/Q5139990","display_name":"Code (set theory)","level":3,"score":0.36239999532699585},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.3571999967098236},{"id":"https://openalex.org/C51823790","wikidata":"https://www.wikidata.org/wiki/Q504353","display_name":"Greedy algorithm","level":2,"score":0.3336000144481659},{"id":"https://openalex.org/C11577676","wikidata":"https://www.wikidata.org/wiki/Q134237","display_name":"Square root","level":2,"score":0.3327000141143799},{"id":"https://openalex.org/C171752962","wikidata":"https://www.wikidata.org/wiki/Q255166","display_name":"Kullback\u2013Leibler divergence","level":2,"score":0.3269999921321869},{"id":"https://openalex.org/C107673813","wikidata":"https://www.wikidata.org/wiki/Q812534","display_name":"Bayesian probability","level":2,"score":0.3012000024318695},{"id":"https://openalex.org/C78548338","wikidata":"https://www.wikidata.org/wiki/Q2493","display_name":"Data compression","level":2,"score":0.29600000381469727},{"id":"https://openalex.org/C2780813799","wikidata":"https://www.wikidata.org/wiki/Q3274237","display_name":"Zero (linguistics)","level":2,"score":0.29010000824928284},{"id":"https://openalex.org/C25797200","wikidata":"https://www.wikidata.org/wiki/Q828137","display_name":"Compression ratio","level":3,"score":0.2890999913215637},{"id":"https://openalex.org/C116714509","wikidata":"https://www.wikidata.org/wiki/Q32278","display_name":"Truncated mean","level":3,"score":0.28859999775886536},{"id":"https://openalex.org/C163175372","wikidata":"https://www.wikidata.org/wiki/Q3339222","display_name":"Linear model","level":2,"score":0.2793000042438507},{"id":"https://openalex.org/C77553402","wikidata":"https://www.wikidata.org/wiki/Q13222579","display_name":"Upper and lower bounds","level":2,"score":0.2770000100135803},{"id":"https://openalex.org/C87007009","wikidata":"https://www.wikidata.org/wiki/Q210832","display_name":"Statistical hypothesis testing","level":2,"score":0.2678999900817871},{"id":"https://openalex.org/C177918212","wikidata":"https://www.wikidata.org/wiki/Q803623","display_name":"Perturbation (astronomy)","level":2,"score":0.26649999618530273},{"id":"https://openalex.org/C40969351","wikidata":"https://www.wikidata.org/wiki/Q3516228","display_name":"Word error rate","level":2,"score":0.26440000534057617},{"id":"https://openalex.org/C149782125","wikidata":"https://www.wikidata.org/wiki/Q160039","display_name":"Econometrics","level":1,"score":0.2574000060558319},{"id":"https://openalex.org/C115051666","wikidata":"https://www.wikidata.org/wiki/Q6522493","display_name":"Ranging","level":2,"score":0.2572999894618988},{"id":"https://openalex.org/C196083921","wikidata":"https://www.wikidata.org/wiki/Q7915758","display_name":"Variance (accounting)","level":2,"score":0.25600001215934753},{"id":"https://openalex.org/C48921125","wikidata":"https://www.wikidata.org/wiki/Q10861030","display_name":"Linear regression","level":2,"score":0.25540000200271606},{"id":"https://openalex.org/C136625980","wikidata":"https://www.wikidata.org/wiki/Q663208","display_name":"Rounding","level":2,"score":0.25369998812675476}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2610.00694","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.00694","pdf_url":"https://arxiv.org/pdf/2610.00694","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.2610.00694","doi":"https://doi.org/10.48550/arxiv.2610.00694","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2610.00694","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:2610.00694","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.00694","pdf_url":"https://arxiv.org/pdf/2610.00694","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":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7216460569.pdf","grobid_xml":"https://content.openalex.org/works/W7216460569.grobid-xml"},"referenced_works_count":47,"referenced_works":["https://openalex.org/W1845051632","https://openalex.org/W2525332836","https://openalex.org/W2626778328","https://openalex.org/W2890152612","https://openalex.org/W2982041622","https://openalex.org/W3042879175","https://openalex.org/W3083410900","https://openalex.org/W4287254789","https://openalex.org/W4307934016","https://openalex.org/W4309591680","https://openalex.org/W4310561894","https://openalex.org/W4312056202","https://openalex.org/W4313484599","https://openalex.org/W4379260375","https://openalex.org/W4380714727","https://openalex.org/W4381586827","https://openalex.org/W4386231419","https://openalex.org/W4387322997","https://openalex.org/W4387561528","https://openalex.org/W4387635845","https://openalex.org/W4390833798","https://openalex.org/W4391335474","https://openalex.org/W4391375196","https://openalex.org/W4392340571","https://openalex.org/W4392490075","https://openalex.org/W4393248014","https://openalex.org/W4395112660","https://openalex.org/W4395474395","https://openalex.org/W4401306886","https://openalex.org/W4405887239","https://openalex.org/W4406032137","https://openalex.org/W4407683611","https://openalex.org/W4412887764","https://openalex.org/W4412889927","https://openalex.org/W4414939355","https://openalex.org/W4415795507","https://openalex.org/W4415796210","https://openalex.org/W4415799217","https://openalex.org/W6947991961","https://openalex.org/W6966825443","https://openalex.org/W7133216168","https://openalex.org/W7133231030","https://openalex.org/W7133245307","https://openalex.org/W7165419702","https://openalex.org/W7167595835","https://openalex.org/W7167932887","https://openalex.org/W7212028571"],"related_works":["https://openalex.org/W4415795507","https://openalex.org/W7155617954","https://openalex.org/W7142807256","https://openalex.org/W7143510021","https://openalex.org/W7197000284","https://openalex.org/W154587749","https://openalex.org/W7203668531","https://openalex.org/W7172501454","https://openalex.org/W4412944786","https://openalex.org/W4388444351"],"abstract_inverted_index":{"Compression":[0],"reports":[1,131],"summarize":[2],"how":[3,31],"far":[4],"a":[5,16,19,81,100,136,182,248],"compressed":[6,49],"language":[7],"model":[8],"moved":[9],"from":[10],"the":[11,24,66,70,115,157,161,179,186,255],"dense":[12,25],"one,":[13],"usually":[14],"by":[15,151],"KL":[17,90,110,155,197,259],"divergence;":[18],"deployment":[20],"that":[21,39,102,163,258],"relies":[22],"on":[23,57,142,181,207,228],"model's":[26],"outputs":[27],"needs":[28],"to":[29,160],"know":[30],"many":[32],"of":[33,53,138,169,202,221,252],"its":[34,96,217],"decisions":[35,166],"changed.":[36],"We":[37],"show":[38],"total":[40,78,171,237],"variation,":[41],"not":[42],"KL,":[43],"answers":[44],"this":[45],"directly.":[46],"Across":[47],"802":[48],"and":[50,60,99,107,145,206],"perturbed":[51],"copies":[52],"19":[54],"open":[55],"models":[56,106,144,192,210],"five":[58],"corpora":[59,146],"nine":[61],"mechanically":[62],"unrelated":[63],"perturbation":[64],"families,":[65],"rate":[67],"at":[68,80,152],"which":[69],"arg-max":[71],"token":[72],"changes":[73,164],"(the":[74],"\\emph{flip":[75],"rate})":[76],"tracks":[77],"variation":[79,172,238],"ratio":[82,187],"with":[83,86,211,247],"median":[84],"$1.05$,":[85],"no":[87],"fitted":[88],"constant.":[89],"converts":[91],"into":[92],"flips":[93],"only":[94],"through":[95],"square":[97],"root":[98,116],"factor":[101],"varies":[103],"fourfold":[104],"across":[105],"corpora,":[108],"because":[109],"averages":[111],"over":[112],"tokens":[113],"before":[114],"is":[117],"taken;":[118],"first-order":[119],"statistics":[120],"averaged":[121],"per":[122],"token,":[123],"such":[124],"as":[125],"Hellinger":[126],"distance,":[127],"avoid":[128],"this,":[129],"but":[130,224],"rarely":[132],"give":[133],"them.":[134],"As":[135],"result,":[137],"two":[139,231],"compressors":[140],"reported":[141],"different":[143],"whose":[147],"flip":[148],"rates":[149],"differ":[150],"least":[153],"$10%$,":[154],"assigns":[156],"smaller":[158],"divergence":[159],"one":[162,227],"more":[165],"in":[167,173,216],"$11%$":[168],"cases,":[170],"$1%$.":[174],"Two":[175],"pre-registered":[176],"tests":[177],"mark":[178],"limits:":[180],"held-out":[183],"code":[184,229],"corpus":[185],"held":[188],"for":[189,200,219,230],"all":[190],"eight":[191],"while":[193],"three":[194,208],"predictions":[195],"about":[196],"each":[198],"failed":[199],"half":[201],"them":[203],"or":[204],"more,":[205],"new":[209],"real":[212],"kernels":[213],"it":[214],"stayed":[215],"band":[218],"37":[220],"38":[222],"checkpoints":[223],"fell":[225],"below":[226],"models.":[232],"In":[233],"vLLM":[234],"speculative":[235],"decoding,":[236],"measured":[239],"under":[240],"teacher":[241],"forcing":[242],"predicts":[243],"greedy":[244],"draft":[245],"acceptance":[246],"mean":[249],"relative":[250],"error":[251],"$1.1$--$2.4%$,":[253],"without":[254],"task-specific":[256],"calibration":[257],"needs.":[260]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-10-03T00:00:00"}