{"id":"https://openalex.org/W7204805597","doi":"https://doi.org/10.18653/2026.emnlp-main.581","title":"SABER: Stability-Aware Early Exit for LLM Reasoning via Adversarial Branch Probing","display_name":"SABER: Stability-Aware Early Exit for LLM Reasoning via Adversarial Branch Probing","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7204805597","doi":"https://doi.org/10.18653/2026.emnlp-main.581"},"language":"en","primary_location":{"id":"doi:10.18653/2026.emnlp-main.581","doi":"https://doi.org/10.18653/2026.emnlp-main.581","is_oa":true,"landing_page_url":"https://doi.org/10.18653/2026.emnlp-main.581","pdf_url":"https://aclanthology.org/2026.emnlp-main.581.pdf","source":{"id":"https://openalex.org/S4363608991","display_name":"Conference on Empirical Methods in Natural Language Processing (EMNLP)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},"volume":null,"type":"conference-paper","indexed_in":["arxiv","crossref","datacite"],"open_access":{"is_oa":true,"oa_status":"bronze","oa_url":"https://aclanthology.org/2026.emnlp-main.581.pdf","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5020738195","display_name":"Wanli Cheng","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I3923682","display_name":"Soochow University","ror":"https://ror.org/05t8y2r12","country_code":"CN","type":"education","lineage":["https://openalex.org/I3923682"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Wanli Cheng","raw_affiliation_strings":["Soochow University , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Soochow University , China","institution_ids":["https://openalex.org/I3923682"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5124028601","display_name":"Haiya Xiang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I3923682","display_name":"Soochow University","ror":"https://ror.org/05t8y2r12","country_code":"CN","type":"education","lineage":["https://openalex.org/I3923682"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Haiya Xiang","raw_affiliation_strings":["Soochow University , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Soochow University , China","institution_ids":["https://openalex.org/I3923682"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138281988","display_name":"Juntao Li","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I3923682","display_name":"Soochow University","ror":"https://ror.org/05t8y2r12","country_code":"CN","type":"education","lineage":["https://openalex.org/I3923682"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Juntao Li","raw_affiliation_strings":["Soochow University , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Soochow University , China","institution_ids":["https://openalex.org/I3923682"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100617350","display_name":"Hongling Wang","orcid":"https://orcid.org/0000-0002-7259-9417","observed_orcids":["https://orcid.org/0000-0002-7259-9417","https://orcid.org/0000-0002-4816-4771"]},"institutions":[{"id":"https://openalex.org/I3923682","display_name":"Soochow University","ror":"https://ror.org/05t8y2r12","country_code":"CN","type":"education","lineage":["https://openalex.org/I3923682"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Hongling Wang","raw_affiliation_strings":["Soochow University , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Soochow University , China","institution_ids":["https://openalex.org/I3923682"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5134169149","display_name":"Wenliang Chen","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I3923682","display_name":"Soochow University","ror":"https://ror.org/05t8y2r12","country_code":"CN","type":"education","lineage":["https://openalex.org/I3923682"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Wenliang Chen","raw_affiliation_strings":["Soochow University , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Soochow University , China","institution_ids":["https://openalex.org/I3923682"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I3923682"],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.69289921,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"13294","last_page":"13313"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.6015999913215637,"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":0.6015999913215637,"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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.1542000025510788,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.10270000249147415,"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/early-exit","display_name":"early exit","score":0.9449999928474426},{"id":"https://openalex.org/keywords/large-reasoning-models","display_name":"large reasoning models","score":0.7839999794960022},{"id":"https://openalex.org/keywords/reasoning-stability","display_name":"reasoning stability","score":0.7039999961853027},{"id":"https://openalex.org/keywords/semantic-perturbation","display_name":"semantic perturbation","score":0.8309999704360962},{"id":"https://openalex.org/keywords/inference-efficiency","display_name":"inference efficiency","score":0.6129999756813049},{"id":"https://openalex.org/keywords/long-chain-reasoning","display_name":"long-chain reasoning","score":0.8029999732971191},{"id":"https://openalex.org/keywords/training-free-framework","display_name":"training-free framework","score":0.6039999723434448}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6764000058174133},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6452999711036682},{"id":"https://openalex.org/C37736160","wikidata":"https://www.wikidata.org/wiki/Q1801315","display_name":"Adversarial system","level":2,"score":0.6295999884605408},{"id":"https://openalex.org/C89288958","wikidata":"https://www.wikidata.org/wiki/Q7301504","display_name":"Reasoning system","level":2,"score":0.6105999946594238},{"id":"https://openalex.org/C86827895","wikidata":"https://www.wikidata.org/wiki/Q7098582","display_name":"Opportunistic reasoning","level":4,"score":0.5370000004768372},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5246000289916992},{"id":"https://openalex.org/C83725634","wikidata":"https://www.wikidata.org/wiki/Q7268699","display_name":"Qualitative reasoning","level":2,"score":0.492000013589859},{"id":"https://openalex.org/C97364631","wikidata":"https://www.wikidata.org/wiki/Q484284","display_name":"Deductive reasoning","level":2,"score":0.4779999852180481},{"id":"https://openalex.org/C159032336","wikidata":"https://www.wikidata.org/wiki/Q2488768","display_name":"Non-monotonic logic","level":2,"score":0.47360000014305115},{"id":"https://openalex.org/C103057564","wikidata":"https://www.wikidata.org/wiki/Q4751139","display_name":"Analytic reasoning","level":3,"score":0.4713999927043915},{"id":"https://openalex.org/C20162079","wikidata":"https://www.wikidata.org/wiki/Q1151406","display_name":"Case-based reasoning","level":2,"score":0.4214000105857849},{"id":"https://openalex.org/C106301342","wikidata":"https://www.wikidata.org/wiki/Q4117933","display_name":"Entropy (arrow of time)","level":2,"score":0.4115000069141388},{"id":"https://openalex.org/C115086926","wikidata":"https://www.wikidata.org/wiki/Q17004651","display_name":"Causal reasoning","level":3,"score":0.3962000012397766},{"id":"https://openalex.org/C37335422","wikidata":"https://www.wikidata.org/wiki/Q6888134","display_name":"Model-based reasoning","level":3,"score":0.3905999958515167},{"id":"https://openalex.org/C13662910","wikidata":"https://www.wikidata.org/wiki/Q193139","display_name":"Trajectory","level":2,"score":0.3831000030040741},{"id":"https://openalex.org/C195344581","wikidata":"https://www.wikidata.org/wiki/Q2555318","display_name":"Automated reasoning","level":2,"score":0.3813000023365021},{"id":"https://openalex.org/C2777508537","wikidata":"https://www.wikidata.org/wiki/Q7936620","display_name":"Visual reasoning","level":2,"score":0.3693000078201294},{"id":"https://openalex.org/C183521366","wikidata":"https://www.wikidata.org/wiki/Q7256422","display_name":"Psychology of reasoning","level":4,"score":0.36230000853538513},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.3582000136375427},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3296000063419342},{"id":"https://openalex.org/C184337299","wikidata":"https://www.wikidata.org/wiki/Q1437428","display_name":"Semantics (computer science)","level":2,"score":0.31380000710487366},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.2994000017642975},{"id":"https://openalex.org/C36964233","wikidata":"https://www.wikidata.org/wiki/Q7920942","display_name":"Verbal reasoning","level":3,"score":0.2980000078678131},{"id":"https://openalex.org/C2780586882","wikidata":"https://www.wikidata.org/wiki/Q7520643","display_name":"Simple (philosophy)","level":2,"score":0.28130000829696655},{"id":"https://openalex.org/C9679016","wikidata":"https://www.wikidata.org/wiki/Q1417473","display_name":"Principle of maximum entropy","level":2,"score":0.2745000123977661},{"id":"https://openalex.org/C193221554","wikidata":"https://www.wikidata.org/wiki/Q5153664","display_name":"Commonsense reasoning","level":2,"score":0.27390000224113464},{"id":"https://openalex.org/C2776325391","wikidata":"https://www.wikidata.org/wiki/Q6917865","display_name":"Motivated reasoning","level":3,"score":0.2655999958515167},{"id":"https://openalex.org/C148220186","wikidata":"https://www.wikidata.org/wiki/Q7111912","display_name":"Outcome (game theory)","level":2,"score":0.259799987077713}],"mesh":[],"locations_count":3,"locations":[{"id":"doi:10.18653/2026.emnlp-main.581","doi":"https://doi.org/10.18653/2026.emnlp-main.581","is_oa":true,"landing_page_url":"https://doi.org/10.18653/2026.emnlp-main.581","pdf_url":"https://aclanthology.org/2026.emnlp-main.581.pdf","source":{"id":"https://openalex.org/S4363608991","display_name":"Conference on Empirical Methods in Natural Language Processing (EMNLP)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},{"id":"pmh:oai:arXiv.org:2608.27963","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.27963","pdf_url":"https://arxiv.org/pdf/2608.27963","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.2608.27963","doi":"https://doi.org/10.48550/arxiv.2608.27963","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2608.27963","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":"doi:10.18653/2026.emnlp-main.581","doi":"https://doi.org/10.18653/2026.emnlp-main.581","is_oa":true,"landing_page_url":"https://doi.org/10.18653/2026.emnlp-main.581","pdf_url":"https://aclanthology.org/2026.emnlp-main.581.pdf","source":{"id":"https://openalex.org/S4363608991","display_name":"Conference on Empirical Methods in Natural Language Processing (EMNLP)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320321001","display_name":"National Natural Science Foundation of China","ror":"https://ror.org/01h0zpd94"},{"id":"https://openalex.org/F4320324720","display_name":"Soochow University","ror":"https://ror.org/05t8y2r12"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7204805597.pdf","grobid_xml":"https://content.openalex.org/works/W7204805597.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7161709774","https://openalex.org/W7128616912","https://openalex.org/W7127542656","https://openalex.org/W7172244577","https://openalex.org/W7138912107","https://openalex.org/W4414633696","https://openalex.org/W4415020792","https://openalex.org/W7153671713","https://openalex.org/W4414804287","https://openalex.org/W6979230633"],"abstract_inverted_index":{"Large":[0],"Reasoning":[1],"Models":[2],"(LRMs)":[3],"achieve":[4],"strong":[5],"reasoning":[6,10,19,22,41,84,120,129],"capabilities,":[7],"yet":[8,78],"long-chain":[9],"becomes":[11],"inefficient":[12],"once":[13],"the":[14,104],"intermediate":[15,83],"answer":[16],"stabilizes":[17],"across":[18,109,118],"steps:":[20],"additional":[21],"yields":[23],"little":[24],"marginal":[25],"benefit":[26],"while":[27,43,136],"incurring":[28],"substantial":[29],"inference":[30],"cost.Existing":[31],"early-exit":[32],"methods":[33],"based":[34],"on":[35,47,134],"confidence":[36],"or":[37],"entropy":[38],"poorly":[39],"capture":[40],"stability,":[42],"consistency-based":[44],"approaches":[45],"rely":[46],"multi-step":[48],"trajectory":[49,102],"agreement,":[50],"requiring":[51],"sequential":[52],"evaluations":[53],"that":[54,126],"delay":[55],"exit.To":[56],"better":[57],"balance":[58],"efficiency":[59],"and":[60,90,122],"reliability,":[61],"we":[62],"propose":[63],"SABER,":[64],"a":[65],"training-free":[66],"framework":[67],"for":[68],"stabilityaware":[69],"early":[70],"exit":[71],"via":[72],"adversarial":[73,88],"branch":[74],"probing.SABER":[75],"constructs":[76],"simple":[77],"effective":[79],"semantic":[80],"perturbations":[81],"around":[82],"states":[85],"to":[86,94],"form":[87],"branches,":[89,110],"applies":[91],"lightweight":[92],"probing":[93],"estimate":[95],"their":[96],"likely":[97],"final":[98],"outcomes":[99,106],"without":[100],"full":[101],"rollouts.When":[103],"probed":[105],"remain":[107],"consistent":[108],"SABER":[111,127],"exits":[112],"early;":[113],"otherwise,":[114],"it":[115],"continues":[116],"reasoning.Experiments":[117],"multiple":[119],"benchmarks":[121],"model":[123],"architectures":[124],"show":[125],"reduces":[128],"token":[130],"consumption":[131],"by":[132],"30.2%-39.8%":[133],"average":[135],"maintaining":[137],"competitive":[138],"accuracy":[139],"with":[140],"full-length":[141],"reasoning.1":[142]},"counts_by_year":[],"updated_date":"2026-10-31T12:09:11.096373","created_date":"2026-09-01T00:00:00"}