{"id":"https://openalex.org/W7220375222","doi":"https://doi.org/10.48550/arxiv.2610.02344","title":"Drive vs. Decay: On the Training Dynamics of Joint-Embedding Predictive Architectures","display_name":"Drive vs. Decay: On the Training Dynamics of Joint-Embedding Predictive Architectures","publication_year":2026,"publication_date":"2026-10-01","ids":{"openalex":"https://openalex.org/W7220375222","doi":"https://doi.org/10.48550/arxiv.2610.02344"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2610.02344","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.02344","pdf_url":"https://arxiv.org/pdf/2610.02344","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":"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/2610.02344","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5056401380","display_name":"Jos\u00e9 Lucas De Melo Costa","orcid":"https://orcid.org/0009-0007-5372-9101","observed_orcids":["https://orcid.org/0009-0007-5372-9101"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Costa, Jos\u00e9 Lucas De Melo","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5140171154","display_name":"Seong Woo Ahn","orcid":"https://orcid.org/0009-0003-3499-2503","observed_orcids":["https://orcid.org/0009-0003-3499-2503"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ahn, Seong Woo","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5040162773","display_name":"Fabrice Popineau","orcid":"https://orcid.org/0000-0002-2941-9046","observed_orcids":["https://orcid.org/0000-0002-2941-9046"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Popineau, Fabrice","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5038266268","display_name":"Arpad Rimmel","orcid":"https://orcid.org/0009-0009-7028-3644","observed_orcids":["https://orcid.org/0009-0009-7028-3644"]},"institutions":[{"id":"https://openalex.org/I1294671590","display_name":"Centre National de la Recherche Scientifique","ror":"https://ror.org/02feahw73","country_code":"FR","type":"government","lineage":["https://openalex.org/I1294671590"]},{"id":"https://openalex.org/I277688954","display_name":"Universit\u00e9 Paris-Saclay","ror":"https://ror.org/03xjwb503","country_code":"FR","type":"education","lineage":["https://openalex.org/I277688954"]},{"id":"https://openalex.org/I4210107720","display_name":"CentraleSup\u00e9lec","ror":"https://ror.org/019tcpt25","country_code":"FR","type":"facility","lineage":["https://openalex.org/I277688954","https://openalex.org/I4210107720"]},{"id":"https://openalex.org/I4387152856","display_name":"Laboratoire Interdisciplinaire des Sciences du Num\u00e9rique","ror":"https://ror.org/00rd81916","country_code":"FR","type":"facility","lineage":["https://openalex.org/I1294671590","https://openalex.org/I277688954","https://openalex.org/I4387152856"]}],"countries":["FR"],"is_corresponding":false,"raw_author_name":"Rimmel, Arpad","raw_affiliation_strings":["Universit\u00e9 Paris-Saclay , CNRS , CentraleSup\u00e9lec , LISN , Orsay , France"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Universit\u00e9 Paris-Saclay , CNRS , CentraleSup\u00e9lec , LISN , Orsay , France","institution_ids":["https://openalex.org/I1294671590","https://openalex.org/I277688954","https://openalex.org/I4210107720","https://openalex.org/I4387152856"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5109248825","display_name":"Bich-Li\u00ean Doan","orcid":"https://orcid.org/0009-0001-5853-2486","observed_orcids":["https://orcid.org/0009-0001-5853-2486"]},"institutions":[{"id":"https://openalex.org/I1294671590","display_name":"Centre National de la Recherche Scientifique","ror":"https://ror.org/02feahw73","country_code":"FR","type":"government","lineage":["https://openalex.org/I1294671590"]},{"id":"https://openalex.org/I277688954","display_name":"Universit\u00e9 Paris-Saclay","ror":"https://ror.org/03xjwb503","country_code":"FR","type":"education","lineage":["https://openalex.org/I277688954"]},{"id":"https://openalex.org/I4210107720","display_name":"CentraleSup\u00e9lec","ror":"https://ror.org/019tcpt25","country_code":"FR","type":"facility","lineage":["https://openalex.org/I277688954","https://openalex.org/I4210107720"]},{"id":"https://openalex.org/I4387152856","display_name":"Laboratoire Interdisciplinaire des Sciences du Num\u00e9rique","ror":"https://ror.org/00rd81916","country_code":"FR","type":"facility","lineage":["https://openalex.org/I1294671590","https://openalex.org/I277688954","https://openalex.org/I4387152856"]}],"countries":["FR"],"is_corresponding":false,"raw_author_name":"Doan, Bich-Li\u00ean","raw_affiliation_strings":["Universit\u00e9 Paris-Saclay , CNRS , CentraleSup\u00e9lec , LISN , Orsay , France"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Universit\u00e9 Paris-Saclay , CNRS , CentraleSup\u00e9lec , LISN , Orsay , France","institution_ids":["https://openalex.org/I1294671590","https://openalex.org/I277688954","https://openalex.org/I4210107720","https://openalex.org/I4387152856"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":4,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.7523999810218811,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.7523999810218811,"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/T11612","display_name":"Stochastic Gradient Optimization Techniques","score":0.05640000104904175,"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":0.032999999821186066,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/jepa","display_name":"JEPA","score":0.925000011920929},{"id":"https://openalex.org/keywords/representation-collapse","display_name":"representation collapse","score":0.9580000042915344},{"id":"https://openalex.org/keywords/training-dynamics","display_name":"training dynamics","score":0.6359999775886536},{"id":"https://openalex.org/keywords/spectral-decoupling","display_name":"spectral decoupling","score":0.7440000176429749},{"id":"https://openalex.org/keywords/phase-boundaries","display_name":"phase boundaries","score":0.4449999928474426},{"id":"https://openalex.org/keywords/effective-rank","display_name":"effective rank","score":0.5649999976158142},{"id":"https://openalex.org/keywords/attentional-bias","display_name":"attentional bias","score":0.4490000009536743},{"id":"https://openalex.org/keywords/exponentially-weighted-moving-average","display_name":"exponential moving average","score":0.7820000052452087},{"id":"https://openalex.org/keywords/tabular-data","display_name":"tabular data","score":0.703000009059906}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.586899995803833},{"id":"https://openalex.org/C112972136","wikidata":"https://www.wikidata.org/wiki/Q7595718","display_name":"Stability (learning theory)","level":2,"score":0.5047000050544739},{"id":"https://openalex.org/C127705205","wikidata":"https://www.wikidata.org/wiki/Q5748245","display_name":"Heuristics","level":2,"score":0.46869999170303345},{"id":"https://openalex.org/C2776359362","wikidata":"https://www.wikidata.org/wiki/Q2145286","display_name":"Representation (politics)","level":3,"score":0.4526999890804291},{"id":"https://openalex.org/C61445026","wikidata":"https://www.wikidata.org/wiki/Q217608","display_name":"Fixed point","level":2,"score":0.41190001368522644},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3880999982357025},{"id":"https://openalex.org/C164226766","wikidata":"https://www.wikidata.org/wiki/Q7293202","display_name":"Rank (graph theory)","level":2,"score":0.36730000376701355},{"id":"https://openalex.org/C47446073","wikidata":"https://www.wikidata.org/wiki/Q5165890","display_name":"Control theory (sociology)","level":3,"score":0.34220001101493835},{"id":"https://openalex.org/C2776760102","wikidata":"https://www.wikidata.org/wiki/Q5139990","display_name":"Code (set theory)","level":3,"score":0.3409000039100647},{"id":"https://openalex.org/C79379906","wikidata":"https://www.wikidata.org/wiki/Q3174497","display_name":"Dynamical systems theory","level":2,"score":0.3407000005245209},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.3343999981880188},{"id":"https://openalex.org/C28719098","wikidata":"https://www.wikidata.org/wiki/Q44946","display_name":"Point (geometry)","level":2,"score":0.334199994802475},{"id":"https://openalex.org/C129844170","wikidata":"https://www.wikidata.org/wiki/Q41299","display_name":"Quadratic equation","level":2,"score":0.3303999900817871},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.32440000772476196},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.3199000060558319},{"id":"https://openalex.org/C120936955","wikidata":"https://www.wikidata.org/wiki/Q2155640","display_name":"Empirical research","level":2,"score":0.3179999887943268},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.31450000405311584},{"id":"https://openalex.org/C73301696","wikidata":"https://www.wikidata.org/wiki/Q5469984","display_name":"Formalism (music)","level":3,"score":0.30570000410079956},{"id":"https://openalex.org/C2778355321","wikidata":"https://www.wikidata.org/wiki/Q17079427","display_name":"Identity (music)","level":2,"score":0.2874999940395355},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.2800999879837036},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.26980000734329224},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.257099986076355}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2610.02344","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.02344","pdf_url":"https://arxiv.org/pdf/2610.02344","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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2610.02344","doi":"https://doi.org/10.48550/arxiv.2610.02344","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2610.02344","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:2610.02344","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.02344","pdf_url":"https://arxiv.org/pdf/2610.02344","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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320309966","display_name":"\u00c9cole Normale Sup\u00e9rieure","ror":"https://ror.org/05a0dhs15"},{"id":"https://openalex.org/F4320322263","display_name":"Conseil R\u00e9gional, \u00cele-de-France","ror":"https://ror.org/04ry1rt27"},{"id":"https://openalex.org/F4320322892","display_name":"Centre National de la Recherche Scientifique","ror":"https://ror.org/02feahw73"},{"id":"https://openalex.org/F4320326256","display_name":"Grand \u00c9quipement National De Calcul Intensif","ror":"https://ror.org/0010d1q40"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7220375222.pdf","grobid_xml":"https://content.openalex.org/works/W7220375222.grobid-xml"},"referenced_works_count":14,"referenced_works":["https://openalex.org/W1488435683","https://openalex.org/W2046402929","https://openalex.org/W2108598243","https://openalex.org/W4288080166","https://openalex.org/W4386076428","https://openalex.org/W4389157290","https://openalex.org/W4392752752","https://openalex.org/W4405468713","https://openalex.org/W4415314996","https://openalex.org/W4415796683","https://openalex.org/W4415797979","https://openalex.org/W4415799137","https://openalex.org/W4416338927","https://openalex.org/W7133207595"],"related_works":["https://openalex.org/W7167412682","https://openalex.org/W7163158573","https://openalex.org/W7140108208","https://openalex.org/W7138168546","https://openalex.org/W7223097630","https://openalex.org/W7149425171","https://openalex.org/W7150062083","https://openalex.org/W7214709537","https://openalex.org/W7202198548","https://openalex.org/W7138407851"],"abstract_inverted_index":{"Joint-Embedding":[0],"Predictive":[1],"Architectures":[2],"(JEPAs)":[3],"are":[4],"prone":[5],"to":[6,159],"representation":[7],"collapse,":[8],"typically":[9],"mitigated":[10],"through":[11],"empirical":[12,156],"heuristics.":[13,23],"We":[14],"develop":[15],"an":[16,160],"early-training":[17],"stability":[18,54],"theory":[19],"that":[20,79],"unifies":[21,100],"these":[22],"Linearising":[24],"the":[25,31,69,75,128],"coupled":[26],"JEPA":[27],"gradient":[28],"flow":[29],"around":[30],"trivial":[32],"fixed":[33],"point":[34],"reveals":[35],"two":[36],"competing":[37],"effects:":[38],"a":[39,44,52,85,120],"driving":[40],"force":[41],"($\u03b3$)":[42],"and":[43,65,68,105,137,143,151],"decay":[45],"effect":[46],"($\u03c3$).":[47],"Under":[48],"approximate":[49],"spectral":[50],"decoupling,":[51],"per-mode":[53],"ratio":[55],"$\u03bc_i":[56],"=":[57],"\u03b3_i":[58],"/":[59],"\u03c3_i$":[60],"factorises":[61],"into":[62],"independent":[63],"data-side":[64],"predictor-side":[66],"terms":[67],"count":[70],"of":[71,77],"unstable":[72],"modes":[73],"tracks":[74],"rank":[76,136],"representations":[78],"can":[80],"emerge.":[81],"The":[82],"framework":[83,154],"predicts":[84],"phase":[86],"boundary,":[87],"which":[88],"we":[89,117],"confirm":[90],"empirically":[91],"across":[92],"more":[93],"than":[94],"800":[95],"Tabular-JEPA":[96],"configurations.":[97],"It":[98],"also":[99],"predictor":[101,122],"scaling,":[102],"masking":[103],"ratio,":[104],"EMA":[106],"as":[107],"distinct":[108],"mechanisms":[109],"for":[110],"shifting":[111],"$\u03bc$.":[112],"Guided":[113],"by":[114],"this":[115],"analysis,":[116],"introduce":[118],"ResidualPred,":[119],"transformer":[121],"whose":[123],"attention":[124],"is":[125,168],"biased":[126],"toward":[127],"identity":[129],"at":[130,170],"initialisation;":[131],"it":[132],"improves":[133],"both":[134],"effective":[135],"downstream":[138],"accuracy":[139],"on":[140,147],"tabular":[141],"benchmarks":[142],"in":[144],"I-JEPA":[145],"pretraining":[146],"CIFAR-10,":[148],"CIFAR-100,":[149],"STL-10,":[150],"ImageNet.":[152],"Our":[153],"connects":[155],"collapse-avoidance":[157],"heuristics":[158],"explicit":[161],"dynamical":[162],"picture,":[163],"yielding":[164],"theory-driven":[165],"stabilizers.":[166],"Code":[167],"available":[169],"https://github.com/jose-melo/drive-vs-decay.":[171]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-10-06T00:00:00"}