{"id":"https://openalex.org/W7166875692","doi":"https://doi.org/10.18653/v1/2026.findings-acl.734","title":"When Internalization Fails: Finding Better Targets for Reasoning Compression","display_name":"When Internalization Fails: Finding Better Targets for Reasoning Compression","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166875692","doi":"https://doi.org/10.18653/v1/2026.findings-acl.734"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2026.findings-acl.734","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.734","pdf_url":"https://aclanthology.org/2026.findings-acl.734.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2026.findings-acl.734.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5115764016","display_name":"Mourad Heddaya","orcid":null},"institutions":[{"id":"https://openalex.org/I40347166","display_name":"University of Chicago","ror":"https://ror.org/024mw5h28","country_code":"US","type":"education","lineage":["https://openalex.org/I40347166"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Mourad Heddaya","raw_affiliation_strings":["University of Chicago University of Chicago"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Chicago University of Chicago","institution_ids":["https://openalex.org/I40347166"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5041830853","display_name":"Manley Roberts","orcid":null},"institutions":[{"id":"https://openalex.org/I40347166","display_name":"University of Chicago","ror":"https://ror.org/024mw5h28","country_code":"US","type":"education","lineage":["https://openalex.org/I40347166"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Manley Roberts","raw_affiliation_strings":["University of Chicago University of Chicago"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Chicago University of Chicago","institution_ids":["https://openalex.org/I40347166"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5056195220","display_name":"Rohan Wadhawan","orcid":"https://orcid.org/0000-0001-8100-668X"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Rohan Wadhawan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5139827063","display_name":"Chenhao Tan","orcid":null},"institutions":[{"id":"https://openalex.org/I40347166","display_name":"University of Chicago","ror":"https://ror.org/024mw5h28","country_code":"US","type":"education","lineage":["https://openalex.org/I40347166"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Chenhao Tan","raw_affiliation_strings":["University of Chicago University of Chicago"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Chicago University of Chicago","institution_ids":["https://openalex.org/I40347166"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.86413448,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"14935","last_page":"14946"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.0763000026345253,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.0763000026345253,"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/T10028","display_name":"Topic Modeling","score":0.0738999992609024,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.06499999761581421,"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/compression","display_name":"Compression (physics)","score":0.5067999958992004},{"id":"https://openalex.org/keywords/internalization","display_name":"Internalization","score":0.37380000948905945},{"id":"https://openalex.org/keywords/data-compression","display_name":"Data compression","score":0.3382999897003174},{"id":"https://openalex.org/keywords/focus","display_name":"Focus (optics)","score":0.257999986410141}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5307000279426575},{"id":"https://openalex.org/C180016635","wikidata":"https://www.wikidata.org/wiki/Q2712821","display_name":"Compression (physics)","level":2,"score":0.5067999958992004},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.436599999666214},{"id":"https://openalex.org/C139770010","wikidata":"https://www.wikidata.org/wiki/Q1339807","display_name":"Internalization","level":3,"score":0.37380000948905945},{"id":"https://openalex.org/C78548338","wikidata":"https://www.wikidata.org/wiki/Q2493","display_name":"Data compression","level":2,"score":0.3382999897003174},{"id":"https://openalex.org/C31972630","wikidata":"https://www.wikidata.org/wiki/Q844240","display_name":"Computer vision","level":1,"score":0.2996000051498413},{"id":"https://openalex.org/C192209626","wikidata":"https://www.wikidata.org/wiki/Q190909","display_name":"Focus (optics)","level":2,"score":0.257999986410141},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.2402999997138977},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.22349999845027924},{"id":"https://openalex.org/C127413603","wikidata":"https://www.wikidata.org/wiki/Q11023","display_name":"Engineering","level":0,"score":0.22059999406337738}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.findings-acl.734","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.734","pdf_url":"https://aclanthology.org/2026.findings-acl.734.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.findings-acl.734","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.734","pdf_url":"https://aclanthology.org/2026.findings-acl.734.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166875692.pdf","grobid_xml":"https://content.openalex.org/works/W7166875692.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Reasoning":[0],"language":[1],"models":[2,87],"generate":[3],"long":[4,111],"reasoning":[5,59,108],"traces":[6,17,64,109],"that":[7,45,143,158],"increase":[8],"latency":[9],"and":[10,53,67,84,112,114,157],"cost.We":[11],"study":[12],"how":[13],"to":[14,57,154],"shorten":[15,47],"these":[16],"while":[18],"preserving":[19],"accuracy":[20,74],"on":[21,104,117,148],"competition-level":[22,105],"mathematics.In":[23],"a":[24,161],"teacher-student":[25],"distillation":[26,56,163],"setup,":[27],"we":[28,69],"compare":[29],"three":[30],"approaches:":[31],"(i)":[32],"inference-time":[33],"truncation":[34],"after":[35,93,123],"the":[36,48,120,126],"first":[37],"k":[38],"tokens,":[39],"(ii)":[40],"Implicit":[41],"Chain-of-Thought":[42],"(ICoT)-style":[43],"curricula":[44,100],"progressively":[46],"teacher":[49,121],"trace":[50],"during":[51],"training,":[52],"(iii)":[54],"direct":[55],"shorter":[58],"traces.Using":[60],"NUMINAMATH":[61],"1.5":[62],"with":[63,81],"from":[65],"DEEPSEEK-R1":[66],"QWQ-32B,":[68],"distill":[70],"into":[71],"QWEN2.5-7Band":[72],"measure":[73],"against":[75],"total":[76],"tokens":[77],"generated.We":[78],"find:":[79],"(1)":[80],"standard":[82],"SFT":[83],"first-k":[85],"truncation,":[86],"compensate":[88],"by":[89],"generating":[90],"longer":[91],"text":[92,119],"reasoning,":[94,124,156],"undermining":[95],"token":[96,139],"savings;":[97],"(2)":[98],"ICoT-style":[99],"provide":[101],"little":[102],"benefit":[103],"mathematics,":[106],"where":[107],"are":[110],"diverse;":[113],"(3)":[115],"training":[116],"post-think,":[118],"generates":[122],"achieves":[125],"best":[127],"accuracy-efficiency":[128],"trade-off":[129],"among":[130],"all":[131],"shortened":[132],"targets,":[133],"outperforming":[134],"generic":[135],"summaries":[136],"at":[137],"matched":[138],"budgets.These":[140],"results":[141],"show":[142],"curriculum-based":[144],"internalization":[145],"methods":[146],"effective":[147],"simple":[149],"tasks":[150],"do":[151],"not":[152],"transfer":[153],"complex":[155],"post-think":[159],"provides":[160],"better":[162],"target.":[164]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2026-07-02T00:00:00"}