{"id":"https://openalex.org/W7140557603","doi":"https://doi.org/10.48550/arxiv.2603.23701","title":"The Diminishing Returns of Early-Exit Decoding in Modern LLMs","display_name":"The Diminishing Returns of Early-Exit Decoding in Modern LLMs","publication_year":2026,"publication_date":"2026-03-24","ids":{"openalex":"https://openalex.org/W7140557603","doi":"https://doi.org/10.48550/arxiv.2603.23701"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2603.23701","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.23701","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","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.2603.23701","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5127348994","display_name":"Rui Wei","orcid":null},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Wei, Rui","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5130643636","display_name":"Rui Du","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Du, Rui","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5075787386","display_name":"Hanfei Yu","orcid":"https://orcid.org/0000-0001-5790-4981"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yu, Hanfei","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5127177696","display_name":"Devesh Tiwari","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tiwari, Devesh","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5130686195","display_name":"Jian Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Jian","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5130706321","display_name":"Zhaozhuo Xu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xu, Zhaozhuo","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5130630733","display_name":"Hao Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Hao","raw_affiliation_strings":[],"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":7,"corresponding_author_ids":["https://openalex.org/A5127348994"],"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":0.11739999800920486,"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":0.11739999800920486,"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/T11636","display_name":"Artificial Intelligence in Healthcare and Education","score":0.09290000051259995,"subfield":{"id":"https://openalex.org/subfields/2718","display_name":"Health Informatics"},"field":{"id":"https://openalex.org/fields/27","display_name":"Medicine"},"domain":{"id":"https://openalex.org/domains/4","display_name":"Health Sciences"}},{"id":"https://openalex.org/T13910","display_name":"Computational and Text Analysis Methods","score":0.08630000054836273,"subfield":{"id":"https://openalex.org/subfields/3300","display_name":"General Social Sciences"},"field":{"id":"https://openalex.org/fields/33","display_name":"Social Sciences"},"domain":{"id":"https://openalex.org/domains/2","display_name":"Social Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/limiting","display_name":"Limiting","score":0.6689000129699707},{"id":"https://openalex.org/keywords/metric","display_name":"Metric (unit)","score":0.5284000039100647},{"id":"https://openalex.org/keywords/decoding-methods","display_name":"Decoding methods","score":0.5135999917984009},{"id":"https://openalex.org/keywords/benchmark","display_name":"Benchmark (surveying)","score":0.49380001425743103},{"id":"https://openalex.org/keywords/computation","display_name":"Computation","score":0.4058000147342682},{"id":"https://openalex.org/keywords/transformer","display_name":"Transformer","score":0.38510000705718994}],"concepts":[{"id":"https://openalex.org/C188198153","wikidata":"https://www.wikidata.org/wiki/Q1613840","display_name":"Limiting","level":2,"score":0.6689000129699707},{"id":"https://openalex.org/C176217482","wikidata":"https://www.wikidata.org/wiki/Q860554","display_name":"Metric (unit)","level":2,"score":0.5284000039100647},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.5135999917984009},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5115000009536743},{"id":"https://openalex.org/C185798385","wikidata":"https://www.wikidata.org/wiki/Q1161707","display_name":"Benchmark (surveying)","level":2,"score":0.49380001425743103},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.4058000147342682},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.38510000705718994},{"id":"https://openalex.org/C162324750","wikidata":"https://www.wikidata.org/wiki/Q8134","display_name":"Economics","level":0,"score":0.38359999656677246},{"id":"https://openalex.org/C2780898871","wikidata":"https://www.wikidata.org/wiki/Q860554","display_name":"Performance metric","level":2,"score":0.3431999981403351},{"id":"https://openalex.org/C82876162","wikidata":"https://www.wikidata.org/wiki/Q17096504","display_name":"Latency (audio)","level":2,"score":0.32260000705718994},{"id":"https://openalex.org/C2779227376","wikidata":"https://www.wikidata.org/wiki/Q6505497","display_name":"Layer (electronics)","level":2,"score":0.32089999318122864},{"id":"https://openalex.org/C42058472","wikidata":"https://www.wikidata.org/wiki/Q810214","display_name":"Base (topology)","level":2,"score":0.31779998540878296},{"id":"https://openalex.org/C2776760102","wikidata":"https://www.wikidata.org/wiki/Q5139990","display_name":"Code (set theory)","level":3,"score":0.3093000054359436},{"id":"https://openalex.org/C162118730","wikidata":"https://www.wikidata.org/wiki/Q1128453","display_name":"Actuarial science","level":1,"score":0.2768999934196472},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.27300000190734863},{"id":"https://openalex.org/C149782125","wikidata":"https://www.wikidata.org/wiki/Q160039","display_name":"Econometrics","level":1,"score":0.27079999446868896},{"id":"https://openalex.org/C112930515","wikidata":"https://www.wikidata.org/wiki/Q4389547","display_name":"Risk analysis (engineering)","level":1,"score":0.2694999873638153},{"id":"https://openalex.org/C96146094","wikidata":"https://www.wikidata.org/wiki/Q609057","display_name":"Unification","level":2,"score":0.2623000144958496}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2603.23701","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.23701","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"article"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2603.23701","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.23701","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"In":[0,117],"Large":[1],"Language":[2],"Model":[3],"(LLM)":[4],"inference,":[5],"early-exit":[6,41,46,69,80,94,109,140],"refers":[7],"to":[8,62,76,137],"stopping":[9],"computation":[10],"at":[11],"an":[12],"intermediate":[13,53],"layer":[14,37],"once":[15],"the":[16,78],"prediction":[17],"is":[18],"sufficiently":[19],"confident,":[20],"thereby":[21],"reducing":[22],"latency":[23],"and":[24,33,50,70,85,113,129],"cost.":[25],"However,":[26],"recent":[27],"LLMs":[28,49],"adopt":[29],"improved":[30],"pretraining":[31],"recipes":[32],"architectures":[34],"that":[35,103],"reduce":[36],"redundancy,":[38],"potentially":[39],"limiting":[40],"opportunities.":[42],"We":[43,58,100],"re-evaluate":[44],"layer-wise":[45],"in":[47,93],"modern":[48],"analyze":[51],"how":[52],"representations":[54],"evolve":[55],"during":[56],"training.":[57],"introduce":[59],"a":[60,64,72,90],"metric":[61],"quantify":[63],"model's":[65],"intrinsic":[66],"suitability":[67],"for":[68,74],"propose":[71],"benchmark":[73],"researchers":[75],"explore":[77],"potential":[79,110],"benefits":[81],"on":[82],"different":[83],"models":[84,132],"workloads.":[86],"Our":[87],"results":[88],"show":[89],"diminishing":[91],"trend":[92],"effectiveness":[95],"across":[96],"newer":[97],"model":[98],"generations.":[99],"further":[101],"find":[102],"dense":[104],"transformers":[105],"generally":[106],"offer":[107],"greater":[108],"than":[111,125],"Mixture-of-Experts":[112],"State":[114],"Space":[115],"Models.":[116],"addition,":[118],"larger":[119],"models,":[120],"particularly":[121],"those":[122],"with":[123],"more":[124],"20":[126],"billion":[127],"parameters,":[128],"base":[130],"pretrained":[131],"without":[133],"specialized":[134],"tuning":[135],"tend":[136],"exhibit":[138],"higher":[139],"potential.":[141]},"counts_by_year":[],"updated_date":"2026-03-27T06:05:27.210665","created_date":"2026-03-27T00:00:00"}