{"id":"https://openalex.org/W7214788975","doi":"https://doi.org/10.48550/arxiv.2609.32902","title":"Linger and Lose: Knowledge Collapse in Low-Bit Language Models","display_name":"Linger and Lose: Knowledge Collapse in Low-Bit Language Models","publication_year":2026,"publication_date":"2026-09-26","ids":{"openalex":"https://openalex.org/W7214788975","doi":"https://doi.org/10.48550/arxiv.2609.32902"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2609.32902","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.32902","pdf_url":"https://arxiv.org/pdf/2609.32902","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/2609.32902","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5139981756","display_name":"Prashanna Mani Paudel","orcid":"https://orcid.org/0009-0006-5139-0936","observed_orcids":["https://orcid.org/0009-0006-5139-0936"]},"institutions":[{"id":"https://openalex.org/I12834331","display_name":"University of Wyoming","ror":"https://ror.org/01485tq96","country_code":"US","type":"education","lineage":["https://openalex.org/I12834331"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Paudel, Prashanna Mani","raw_affiliation_strings":["Geometric Intelligence Research Lab , Dept. of Electrical Engineering and Computer Science , University of Wyoming , USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Geometric Intelligence Research Lab , Dept. of Electrical Engineering and Computer Science , University of Wyoming , USA","institution_ids":["https://openalex.org/I12834331"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5014163696","display_name":"Shivanand Venkanna Sheshappanavar","orcid":"https://orcid.org/0000-0003-4039-2910","observed_orcids":["https://orcid.org/0000-0003-4039-2910"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Sheshappanavar, Shivanand Venkanna","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"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/T10036","display_name":"Advanced Neural Network Applications","score":0.7128999829292297,"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"}},"topics":[{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.7128999829292297,"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/T10028","display_name":"Topic Modeling","score":0.11990000307559967,"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":0.0340999998152256,"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/ternary-weights","display_name":"ternary weights","score":0.7570000290870667},{"id":"https://openalex.org/keywords/knowledge-capacity","display_name":"knowledge capacity","score":0.8740000128746033},{"id":"https://openalex.org/keywords/learning-rate-schedules","display_name":"learning rate scheduling","score":0.5529999732971191},{"id":"https://openalex.org/keywords/training-from-scratch","display_name":"training from scratch","score":0.531000018119812},{"id":"https://openalex.org/keywords/post-training-quantization","display_name":"post-training quantization","score":0.7480000257492065},{"id":"https://openalex.org/keywords/gpt-2","display_name":"GPT-2","score":0.5849999785423279},{"id":"https://openalex.org/keywords/perplexity","display_name":"perplexity","score":0.37700000405311584}],"concepts":[{"id":"https://openalex.org/C100279451","wikidata":"https://www.wikidata.org/wiki/Q372193","display_name":"Perplexity","level":3,"score":0.8159999847412109},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.7114999890327454},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.46560001373291016},{"id":"https://openalex.org/C2776291640","wikidata":"https://www.wikidata.org/wiki/Q2912517","display_name":"Value (mathematics)","level":2,"score":0.4219000041484833},{"id":"https://openalex.org/C2780009758","wikidata":"https://www.wikidata.org/wiki/Q6804172","display_name":"Measure (data warehouse)","level":2,"score":0.39410001039505005},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.38339999318122864},{"id":"https://openalex.org/C151637689","wikidata":"https://www.wikidata.org/wiki/Q5318064","display_name":"Dwell time","level":2,"score":0.3580000102519989},{"id":"https://openalex.org/C68387754","wikidata":"https://www.wikidata.org/wiki/Q7271585","display_name":"Schedule","level":2,"score":0.34630000591278076},{"id":"https://openalex.org/C108757681","wikidata":"https://www.wikidata.org/wiki/Q2773912","display_name":"Bigram","level":3,"score":0.3434999883174896},{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.32440000772476196},{"id":"https://openalex.org/C2778112365","wikidata":"https://www.wikidata.org/wiki/Q3511065","display_name":"Sequence (biology)","level":2,"score":0.320499986410141},{"id":"https://openalex.org/C149782125","wikidata":"https://www.wikidata.org/wiki/Q160039","display_name":"Econometrics","level":1,"score":0.31769999861717224},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.3059000074863434},{"id":"https://openalex.org/C105795698","wikidata":"https://www.wikidata.org/wiki/Q12483","display_name":"Statistics","level":1,"score":0.29679998755455017},{"id":"https://openalex.org/C77553402","wikidata":"https://www.wikidata.org/wiki/Q13222579","display_name":"Upper and lower bounds","level":2,"score":0.2939000129699707},{"id":"https://openalex.org/C141513077","wikidata":"https://www.wikidata.org/wiki/Q378542","display_name":"Independent and identically distributed random variables","level":3,"score":0.2831999957561493},{"id":"https://openalex.org/C114289077","wikidata":"https://www.wikidata.org/wiki/Q3284399","display_name":"Statistical model","level":2,"score":0.27399998903274536},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.2728999853134155},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.2694000005722046},{"id":"https://openalex.org/C64452783","wikidata":"https://www.wikidata.org/wiki/Q1524945","display_name":"Ternary operation","level":2,"score":0.266400009393692},{"id":"https://openalex.org/C122507166","wikidata":"https://www.wikidata.org/wiki/Q628906","display_name":"Crossover","level":2,"score":0.26589998602867126},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.25780001282691956},{"id":"https://openalex.org/C2781007418","wikidata":"https://www.wikidata.org/wiki/Q1456934","display_name":"Capacity planning","level":2,"score":0.2547999918460846}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2609.32902","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.32902","pdf_url":"https://arxiv.org/pdf/2609.32902","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.2609.32902","doi":"https://doi.org/10.48550/arxiv.2609.32902","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.32902","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2609.32902","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.32902","pdf_url":"https://arxiv.org/pdf/2609.32902","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":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7214788975.pdf","grobid_xml":"https://content.openalex.org/works/W7214788975.grobid-xml"},"referenced_works_count":19,"referenced_works":["https://openalex.org/W2001570872","https://openalex.org/W2030450972","https://openalex.org/W4392271821","https://openalex.org/W4394709684","https://openalex.org/W4401768430","https://openalex.org/W4403443919","https://openalex.org/W4404404376","https://openalex.org/W4410356599","https://openalex.org/W4414803876","https://openalex.org/W4414982525","https://openalex.org/W4415332740","https://openalex.org/W4415796834","https://openalex.org/W7153261982","https://openalex.org/W7154589510","https://openalex.org/W7160271086","https://openalex.org/W7160961825","https://openalex.org/W7162404810","https://openalex.org/W7163333518","https://openalex.org/W7196965917"],"related_works":["https://openalex.org/W7204909273","https://openalex.org/W7167591956","https://openalex.org/W7167622706","https://openalex.org/W7155528163","https://openalex.org/W7128491674","https://openalex.org/W7207714806","https://openalex.org/W7167626764","https://openalex.org/W4405173573","https://openalex.org/W7214377235","https://openalex.org/W7167633132"],"abstract_inverted_index":{"Training":[0],"language":[1],"models":[2,56,73,105],"with":[3,49,59,89,128,202],"ternary":[4,72,207],"weights":[5,170],"is":[6,181],"commonly":[7],"judged":[8],"by":[9,95,209,256],"loss":[10],"and":[11,109,140,214],"downstream":[12],"accuracy,":[13],"which":[14,195],"record":[15],"only":[16,96],"a":[17,31,121,131,136,143,162,203],"modest":[18],"cost":[19],"relative":[20],"to":[21,61,98,142],"full":[22],"precision.":[23,224],"We":[24,35,53,115,150,183],"show":[25],"that":[26,185],"these":[27,104],"metrics":[28],"can":[29,166],"conceal":[30],"much":[32],"larger":[33,221],"failure.":[34],"instead":[36],"measure":[37],"knowledge":[38,118],"capacity,":[39],"the":[40,68,153,156,164,178,187,191,226],"factual":[41],"bits":[42],"stored":[43],"per":[44],"parameter,":[45],"on":[46],"synthetic":[47],"biographies":[48],"known":[50],"information":[51],"content.":[52],"train":[54],"GPT-2-style":[55],"from":[57,235],"scratch":[58],"2.5M":[60],"50M":[62],"parameters":[63],"at":[64,161,212,217,222],"five":[65],"precisions.":[66],"Under":[67],"standard":[69],"cosine":[70],"schedule,":[71],"retain":[74],"as":[75,77,120],"little":[76],"6%":[78],"of":[79,113,194],"an":[80],"identically":[81],"trained":[82],"fp16":[83],"model's":[84],"capacity.":[85,149],"The":[86,169,237],"deficit":[87],"widens":[88],"model":[90,133,165,179],"size":[91],"while":[92,174],"perplexity":[93],"rises":[94],"1.4":[97],"1.6":[99],"times.":[100],"Measured":[101],"throughout":[102],"training,":[103],"first":[106],"acquire":[107],"capacity":[108,208,246,258],"then":[110,151],"lose":[111],"most":[112],"it.":[114,198],"identify":[116],"this":[117],"collapse":[119,154],"learning-rate":[122],"dwell":[123],"instability.":[124],"Held":[125],"near":[126],"$1$--$2\\times10^{-4}$":[127],"no":[129,244],"decay,":[130],"pre-collapse":[132],"collapses":[134],"within":[135],"few":[137],"hundred":[138],"exposures,":[139],"returning":[141],"safe":[144],"rate":[145,230],"does":[146],"not":[147],"restore":[148],"locate":[152],"in":[155],"output":[157,227],"head.":[158],"It":[159],"happens":[160],"value":[163,188],"never":[167],"predict.":[168],"there":[171],"grow":[172],"unchecked,":[173],"every":[175],"other":[176],"prediction":[177],"makes":[180],"unchanged.":[182],"find":[184],"making":[186],"predictable":[189],"removes":[190],"collapse,":[192],"regardless":[193],"attribute":[196],"carries":[197],"A":[199],"warmup-stable-decay":[200],"schedule":[201],"10%":[204],"cooldown":[205],"increases":[206],"2.7":[210],"times":[211,216],"25M":[213],"4.3":[215],"50M.":[218],"Gains":[219],"are":[220],"lower":[223],"Cutting":[225],"head's":[228],"learning":[229],"prevents":[231],"failure":[232],"when":[233],"training":[234,255,260],"scratch.":[236],"post-training":[238],"quantization":[239],"methods":[240],"we":[241],"tested":[242],"recover":[243],"measurable":[245],"below":[247],"4":[248],"bits.":[249],"Our":[250],"findings":[251],"suggest":[252],"judging":[253],"low-precision":[254],"retained":[257],"during":[259],"rather":[261],"than":[262],"final":[263],"loss.":[264]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-09-30T00:00:00"}