{"id":"https://openalex.org/W4416248367","doi":"https://doi.org/10.48550/arxiv.2510.19669","title":"DiffAdapt: Difficulty-Adaptive Reasoning for Token-Efficient LLM Inference","display_name":"DiffAdapt: Difficulty-Adaptive Reasoning for Token-Efficient LLM Inference","alternate_titles":[],"publication_year":2026,"publication_date":"2026-04-20","ids":{"openalex":"https://openalex.org/W4416248367","doi":"https://doi.org/10.48550/arxiv.2510.19669"},"language":"en","primary_location":{"id":"pmh:oai:proceedings.iclr.cc:2026/353d4deedc4aac78b24924c485cdb4a3","doi":null,"is_oa":true,"landing_page_url":"https://proceedings.iclr.cc/paper_files/paper/2026/hash/353d4deedc4aac78b24924c485cdb4a3-Abstract-Conference.html","pdf_url":"https://proceedings.iclr.cc/paper_files/paper/2026/file/353d4deedc4aac78b24924c485cdb4a3-Paper-Conference.pdf","source":{"id":"https://openalex.org/S4306419637","display_name":"International Conference on Learning Representations","issn_l":null,"issn":null,"is_oa":true,"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":"International Conference on Learning Representations 2026","raw_type":"info:eu-repo/semantics/conferenceObject"},"parent_works":[],"lineage":["https://openalex.org/W4416248367"],"child_works_count":0,"type":"conference-paper","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"diamond","oa_url":"https://proceedings.iclr.cc/paper_files/paper/2026/file/353d4deedc4aac78b24924c485cdb4a3-Paper-Conference.pdf","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5109152867","display_name":"Xiang Liu","orcid":"https://orcid.org/0009-0003-2639-1804","observed_orcids":["https://orcid.org/0009-0003-2639-1804"]},"institutions":[],"countries":[],"is_corresponding":false,"contributor_roles":[],"raw_author_name":"Liu, Xiang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5043651092","display_name":"Xuming Hu","orcid":"https://orcid.org/0000-0001-6075-4224","observed_orcids":["https://orcid.org/0000-0001-6075-4224"]},"institutions":[],"countries":[],"is_corresponding":false,"contributor_roles":[],"raw_author_name":"Hu, Xuming","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100730785","display_name":"Xiaowen Chu","orcid":"https://orcid.org/0000-0001-9745-4372","observed_orcids":["https://orcid.org/0000-0001-9745-4372"]},"institutions":[],"countries":[],"is_corresponding":false,"contributor_roles":[],"raw_author_name":"Chu, Xiaowen","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5035142405","display_name":"Eunsol Choi","orcid":"https://orcid.org/0000-0003-3607-9104","observed_orcids":["https://orcid.org/0000-0003-3607-9104"]},"institutions":[{"id":"https://openalex.org/I4405261251","display_name":"The Hong Kong University of Science and Technology (Guangzhou)","ror":"https://ror.org/050h0vm43","country_code":"CN","type":"education","lineage":["https://openalex.org/I4405261251"]},{"id":"https://openalex.org/I57206974","display_name":"New York University","ror":"https://ror.org/0190ak572","country_code":"US","type":"education","lineage":["https://openalex.org/I57206974"]}],"countries":["CN","US"],"is_corresponding":false,"contributor_roles":[],"raw_author_name":"Choi, Eunsol","raw_affiliation_strings":["New York University","The Hong Kong University of Science and Technology (Guangzhou ) N"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"New York University","institution_ids":["https://openalex.org/I57206974"]},{"raw_affiliation_string":"The Hong Kong University of Science and Technology (Guangzhou ) N","institution_ids":["https://openalex.org/I4405261251"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":2,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":8.9673,"has_fulltext":true,"cited_by_count":14,"reviewed_by_count":0,"citation_normalized_percentile":{"value":0.9777614,"is_in_top_1_percent":false,"is_in_top_10_percent":true},"cited_by_percentile_year":{"min":98,"max":99},"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_pulled":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.7846999764442444,"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":0.7846999764442444,"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":0.16040000319480896,"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":0.008999999612569809,"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/large-language-model-reasoning","display_name":"large language model reasoning","score":0.675000011920929},{"id":"https://openalex.org/keywords/overthinking","display_name":"overthinking","score":0.6119999885559082},{"id":"https://openalex.org/keywords/reasoning-traces","display_name":"reasoning traces","score":0.6029999852180481},{"id":"https://openalex.org/keywords/token-usage-reduction","display_name":"token usage reduction","score":0.8130000233650208}],"concepts":[{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.7228999733924866},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7077999711036682},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.6777999997138977},{"id":"https://openalex.org/C106301342","wikidata":"https://www.wikidata.org/wiki/Q4117933","display_name":"Entropy (arrow of time)","level":2,"score":0.6074000000953674},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5393999814987183},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.447299987077713},{"id":"https://openalex.org/C2779913896","wikidata":"https://www.wikidata.org/wiki/Q7063001","display_name":"Notice","level":2,"score":0.3774000108242035},{"id":"https://openalex.org/C9679016","wikidata":"https://www.wikidata.org/wiki/Q1417473","display_name":"Principle of maximum entropy","level":2,"score":0.3691999912261963},{"id":"https://openalex.org/C83725634","wikidata":"https://www.wikidata.org/wiki/Q7268699","display_name":"Qualitative reasoning","level":2,"score":0.32659998536109924},{"id":"https://openalex.org/C2777735758","wikidata":"https://www.wikidata.org/wiki/Q817765","display_name":"Path (computing)","level":2,"score":0.2896000146865845},{"id":"https://openalex.org/C4554734","wikidata":"https://www.wikidata.org/wiki/Q593744","display_name":"Knowledge base","level":2,"score":0.28529998660087585},{"id":"https://openalex.org/C89288958","wikidata":"https://www.wikidata.org/wiki/Q7301504","display_name":"Reasoning system","level":2,"score":0.2508000135421753}],"mesh":[],"locations_count":3,"locations":[{"id":"pmh:oai:proceedings.iclr.cc:2026/353d4deedc4aac78b24924c485cdb4a3","doi":null,"is_oa":true,"landing_page_url":"https://proceedings.iclr.cc/paper_files/paper/2026/hash/353d4deedc4aac78b24924c485cdb4a3-Abstract-Conference.html","pdf_url":"https://proceedings.iclr.cc/paper_files/paper/2026/file/353d4deedc4aac78b24924c485cdb4a3-Paper-Conference.pdf","source":{"id":"https://openalex.org/S4306419637","display_name":"International Conference on Learning Representations","issn_l":null,"issn":null,"is_oa":true,"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":"International Conference on Learning Representations 2026","raw_type":"info:eu-repo/semantics/conferenceObject"},{"id":"pmh:oai:arXiv.org:2510.19669","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2510.19669","pdf_url":"https://arxiv.org/pdf/2510.19669","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.2510.19669","doi":"https://doi.org/10.48550/arxiv.2510.19669","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2510.19669","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":"pmh:oai:proceedings.iclr.cc:2026/353d4deedc4aac78b24924c485cdb4a3","doi":null,"is_oa":true,"landing_page_url":"https://proceedings.iclr.cc/paper_files/paper/2026/hash/353d4deedc4aac78b24924c485cdb4a3-Abstract-Conference.html","pdf_url":"https://proceedings.iclr.cc/paper_files/paper/2026/file/353d4deedc4aac78b24924c485cdb4a3-Paper-Conference.pdf","source":{"id":"https://openalex.org/S4306419637","display_name":"International Conference on Learning Representations","issn_l":null,"issn":null,"is_oa":true,"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":"International Conference on Learning Representations 2026","raw_type":"info:eu-repo/semantics/conferenceObject"},"sustainable_development_goals":[],"study_designs":[],"awards":[{"id":"https://openalex.org/G2191774043","display_name":null,"funder_award_id":"62272122","funder_id":"https://openalex.org/F4320321001","funder_display_name":"National Natural Science Foundation of China"},{"id":"https://openalex.org/G4626160587","display_name":"CAREER: Building Interactive Language Systems Navigating Rich Knowledge Sources","funder_award_id":"2443271","funder_id":"https://openalex.org/F4320306076","funder_display_name":"National Science Foundation"},{"id":"https://openalex.org/G695605855","display_name":"Collaborative Research: RI: Medium: Multilingual Long-form QA with Retrieval-Augmented Language Models","funder_award_id":"2521091","funder_id":"https://openalex.org/F4320306076","funder_display_name":"National Science Foundation"}],"funders":[{"id":"https://openalex.org/F4320306076","display_name":"National Science Foundation","ror":"https://ror.org/021nxhr62"},{"id":"https://openalex.org/F4320321001","display_name":"National Natural Science Foundation of China","ror":"https://ror.org/01h0zpd94"},{"id":"https://openalex.org/F4320323537","display_name":"Hong Kong University of Science and Technology","ror":"https://ror.org/00q4vv597"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4416248367.pdf","grobid_xml":"https://content.openalex.org/works/W4416248367.grobid-xml"},"referenced_works_count":43,"referenced_works":["https://openalex.org/W3177813494","https://openalex.org/W4379089709","https://openalex.org/W4387321091","https://openalex.org/W4388514508","https://openalex.org/W4400518576","https://openalex.org/W4402671952","https://openalex.org/W4402684156","https://openalex.org/W4404782610","https://openalex.org/W4407123430","https://openalex.org/W4407185131","https://openalex.org/W4407231762","https://openalex.org/W4407690294","https://openalex.org/W4414580500","https://openalex.org/W4414581695","https://openalex.org/W4414633696","https://openalex.org/W4414859475","https://openalex.org/W4414939355","https://openalex.org/W4415063194","https://openalex.org/W4415109349","https://openalex.org/W4415158403","https://openalex.org/W4415161187","https://openalex.org/W4415186420","https://openalex.org/W4415328803","https://openalex.org/W4415329330","https://openalex.org/W4415796274","https://openalex.org/W4415868133","https://openalex.org/W4416035135","https://openalex.org/W4416036427","https://openalex.org/W4416036482","https://openalex.org/W4416052877","https://openalex.org/W4416056890","https://openalex.org/W4416075286","https://openalex.org/W4416113410","https://openalex.org/W4416649973","https://openalex.org/W4417287872","https://openalex.org/W6929293117","https://openalex.org/W6966994820","https://openalex.org/W6967114859","https://openalex.org/W7133211296","https://openalex.org/W7133224126","https://openalex.org/W7166834929","https://openalex.org/W7166881697","https://openalex.org/W7167559498"],"related_works":["https://openalex.org/W7166885476","https://openalex.org/W7075993915","https://openalex.org/W6929723227","https://openalex.org/W4416434837","https://openalex.org/W7138085980","https://openalex.org/W7169107258","https://openalex.org/W4410355389","https://openalex.org/W7083842473","https://openalex.org/W7158981081","https://openalex.org/W4414581013"],"abstract_inverted_index":{"Recent":[0],"reasoning":[1,44,120],"Large":[2],"Language":[3],"Models":[4],"(LLMs)":[5],"demonstrate":[6],"remarkable":[7],"problem-solving":[8],"abilities":[9],"but":[10,150],"often":[11],"generate":[12],"long":[13],"thinking":[14],"traces":[15],"whose":[16],"utility":[17],"is":[18],"unclear.":[19],"Our":[20,174],"work":[21],"aims":[22],"to":[23,29,87,138,187],"improve":[24],"their":[25,117],"efficiency,":[26],"enabling":[27],"them":[28],"reach":[30],"high":[31,56,62,72],"performance":[32],"without":[33],"overthinking.":[34],"First,":[35],"we":[36,49,80,102],"analyze":[37],"the":[38],"entropy":[39,54,57,65,73,83],"of":[40,127],"token":[41,134,183],"probabilities":[42],"in":[43],"traces.":[45],"Across":[46],"three":[47],"models,":[48],"observe":[50],"a":[51,105,128,151,190],"consistent":[52],"U-shaped":[53],"pattern:":[55],"on":[58,66,74,95,99,116,168],"easy":[59,86,96],"problems":[60,67,76],"despite":[61],"accuracy,":[63],"low":[64],"with":[68],"medium":[69,88],"difficulty,":[70],"and":[71,119,132,171],"hard":[75],"reflecting":[77],"uncertainty.":[78],"Specifically,":[79],"notice":[81],"22--25\\%":[82],"reduction":[84],"from":[85],"difficulty":[89,118],"regions,":[90],"suggesting":[91],"an":[92],"{overthinking}":[93],"phenomenon":[94],"instances.":[97],"Building":[98],"these":[100],"insights,":[101],"introduce":[103],"\\textbf{DiffAdapt},":[104],"lightweight":[106],"framework":[107],"that":[108,154],"selects":[109],"Easy/Normal/Hard":[110],"inference":[111,124],"strategies":[112],"per":[113],"question":[114],"based":[115],"trace":[121],"entropy.":[122],"Each":[123],"strategy":[125],"consists":[126],"fixed":[129],"prompt,":[130],"temperature":[131],"maximum":[133],"length.":[135],"In":[136],"contrast":[137],"existing":[139],"efficiency":[140],"optimization":[141],"methods,":[142],"our":[143,166],"approach":[144],"does":[145],"not":[146],"fine-tune":[147],"base":[148],"LLM":[149],"small":[152],"probe":[153],"classifies":[155],"LLM's":[156],"final":[157],"hidden":[158],"state,":[159],"allowing":[160],"inexpensive":[161],"adaptation.":[162],"We":[163],"comprehensively":[164],"evaluate":[165],"method":[167,175],"five":[169],"models":[170],"eight":[172],"benchmarks.":[173],"achieves":[176],"comparable":[177],"or":[178],"improved":[179],"accuracy":[180],"while":[181],"reducing":[182],"usage":[184],"by":[185],"up":[186],"22.4\\%,":[188],"establishing":[189],"practical":[191],"path":[192],"toward":[193],"compute-efficient":[194],"reasoning.":[195]},"cited_by_api_url":"https://api.openalex.org/works?filter=cites:W4416248367","counts_by_year":[{"year":2026,"cited_by_count":14}],"updated_date":"2026-11-22T12:46:00.618207","created_date":"2025-10-24T00:00:00"}
