{"id":"https://openalex.org/W7133334822","doi":"https://doi.org/10.48550/arxiv.2603.02023","title":"PonderLM-3: Adaptive Token-Wise Pondering with Differentiable Masking","display_name":"PonderLM-3: Adaptive Token-Wise Pondering with Differentiable Masking","publication_year":2026,"publication_date":"2026-03-02","ids":{"openalex":"https://openalex.org/W7133334822","doi":"https://doi.org/10.48550/arxiv.2603.02023"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2603.02023","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.02023","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":"article"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2603.02023","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5128028780","display_name":"He Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, He","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5125566081","display_name":"Feichen Song","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Song, Feichen","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5125504206","display_name":"Boyi Zeng","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zeng, Boyi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5075075278","display_name":"Shixiang Song","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Song, Shixiang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5125893253","display_name":"Zhiqin John Xu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xu, Zhiqin John","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5127907753","display_name":"Ziwei He","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"He, Ziwei","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":null,"display_name":"Lin, Zhouhan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lin, Zhouhan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":7,"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":0.42160001397132874,"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":0.42160001397132874,"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":0.1251000016927719,"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":0.08720000088214874,"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/computation","display_name":"Computation","score":0.7572000026702881},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.7121999859809875},{"id":"https://openalex.org/keywords/differentiable-function","display_name":"Differentiable function","score":0.6847000122070312},{"id":"https://openalex.org/keywords/pruning","display_name":"Pruning","score":0.5472000241279602},{"id":"https://openalex.org/keywords/masking","display_name":"Masking (illustration)","score":0.43689998984336853},{"id":"https://openalex.org/keywords/pareto-optimal","display_name":"Pareto optimal","score":0.40070000290870667},{"id":"https://openalex.org/keywords/matching","display_name":"Matching (statistics)","score":0.3977000117301941}],"concepts":[{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.7572000026702881},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.7121999859809875},{"id":"https://openalex.org/C202615002","wikidata":"https://www.wikidata.org/wiki/Q783507","display_name":"Differentiable function","level":2,"score":0.6847000122070312},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5875999927520752},{"id":"https://openalex.org/C108010975","wikidata":"https://www.wikidata.org/wiki/Q500094","display_name":"Pruning","level":2,"score":0.5472000241279602},{"id":"https://openalex.org/C2777402240","wikidata":"https://www.wikidata.org/wiki/Q6783436","display_name":"Masking (illustration)","level":2,"score":0.43689998984336853},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.40959998965263367},{"id":"https://openalex.org/C2986314615","wikidata":"https://www.wikidata.org/wiki/Q36829","display_name":"Pareto optimal","level":3,"score":0.40070000290870667},{"id":"https://openalex.org/C165064840","wikidata":"https://www.wikidata.org/wiki/Q1321061","display_name":"Matching (statistics)","level":2,"score":0.3977000117301941},{"id":"https://openalex.org/C55439883","wikidata":"https://www.wikidata.org/wiki/Q360812","display_name":"Correctness","level":2,"score":0.38510000705718994},{"id":"https://openalex.org/C137635306","wikidata":"https://www.wikidata.org/wiki/Q182667","display_name":"Pareto principle","level":2,"score":0.3718000054359436},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.35100001096725464},{"id":"https://openalex.org/C99844830","wikidata":"https://www.wikidata.org/wiki/Q102441924","display_name":"Scaling","level":2,"score":0.34040001034736633},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.32589998841285706},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.31470000743865967},{"id":"https://openalex.org/C3770464","wikidata":"https://www.wikidata.org/wiki/Q775963","display_name":"Smoothing","level":2,"score":0.2815000116825104},{"id":"https://openalex.org/C39927690","wikidata":"https://www.wikidata.org/wiki/Q11197","display_name":"Logarithm","level":2,"score":0.2786000072956085},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.27079999446868896},{"id":"https://openalex.org/C3826847","wikidata":"https://www.wikidata.org/wiki/Q188768","display_name":"FLOPS","level":2,"score":0.25200000405311584}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2603.02023","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.02023","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":"article"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2603.02023","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.02023","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":"article"},"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":{"Test-time":[0],"scaling":[1],"has":[2],"shown":[3],"that":[4,40],"allocating":[5],"more":[6,70],"additional":[7,45,60,150,175],"computation":[8,23,46,62,71,151],"at":[9,110,130],"inference":[10,61,132,156,176],"can":[11],"improve":[12],"generation":[13],"quality,":[14],"motivating":[15],"a":[16,33,80,95,105,114],"natural":[17],"follow-up":[18],"question:":[19],"where":[20,181],"should":[21],"this":[22,28,86],"be":[24,179],"spent?":[25],"Building":[26],"on":[27,52],"insight,":[29],"we":[30],"introduce":[31],"PonderLM-3,":[32],"pretraining":[34,100,128],"framework":[35,169],"for":[36,170],"token-wise":[37,171],"adaptive":[38,123,172],"pondering":[39],"learns":[41],"to":[42,141,178],"selectively":[43],"allocate":[44],"under":[47,144],"purely":[48],"self-supervised":[49],"objectives,":[50],"built":[51],"top":[53],"of":[54,149],"the":[55,145],"PonderLM-2":[56,143],"backbone.":[57],"This":[58],"makes":[59],"an":[63,163],"allocatable":[64],"per-token":[65],"resource,":[66],"so":[67],"tokens":[68],"receive":[69],"only":[72],"when":[73],"it":[74,103,125,182],"is":[75,183],"beneficial,":[76],"rather":[77,186],"than":[78,187],"paying":[79],"uniform":[81],"extra":[82],"cost.":[83],"To":[84],"make":[85],"allocation":[87],"learnable":[88],"while":[89,153],"maintaining":[90],"train-inference":[91,167],"consistency,":[92],"PonderLM-3":[93,112,137,161],"injects":[94],"differentiable":[96,165],"attention":[97],"mask":[98],"during":[99],"and":[101,166],"pairs":[102],"with":[104,119],"matching":[106],"hard":[107],"pruning":[108],"rule":[109],"inference.":[111],"defines":[113],"stronger":[115],"Pareto":[116],"frontier:":[117],"compared":[118],"existing":[120],"recursive":[121],"or":[122],"baselines,":[124],"achieves":[126],"lower":[127],"perplexity":[129],"equal":[131],"FLOPs.":[133],"On":[134],"downstream":[135],"benchmarks,":[136],"attains":[138],"comparable":[139],"performance":[140],"fixed-step":[142],"same":[146],"maximum":[147],"number":[148],"steps,":[152],"using":[154],"fewer":[155],"FLOPs":[157],"in":[158],"practice.":[159],"Overall,":[160],"provides":[162],"end-to-end":[164],"consistent":[168],"computation,":[173],"enabling":[174],"compute":[177],"allocated":[180],"most":[184],"useful":[185],"paid":[188],"uniformly":[189],"by":[190],"every":[191],"token.":[192]},"counts_by_year":[],"updated_date":"2026-06-11T09:08:48.828518","created_date":"2026-03-04T00:00:00"}