{"id":"https://openalex.org/W7166827583","doi":"https://doi.org/10.18653/v1/2026.acl-long.990","title":"Step-GRPO: Internalizing Dynamic Early Exit for Efficient Reasoning","display_name":"Step-GRPO: Internalizing Dynamic Early Exit for Efficient Reasoning","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166827583","doi":"https://doi.org/10.18653/v1/2026.acl-long.990"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2026.acl-long.990","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.990","pdf_url":"https://aclanthology.org/2026.acl-long.990.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":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2026.acl-long.990.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5066415914","display_name":"Benteng Chen","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Benteng Chen","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139804313","display_name":"Weida Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Weida Wang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139766949","display_name":"Shufei Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Shufei Zhang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5017033486","display_name":"Mingbao Lin","orcid":"https://orcid.org/0000-0003-1764-1894"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Mingbao Lin","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5139762730","display_name":"Min Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Min Zhang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"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":0.81732099,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"21710","last_page":"21724"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10906","display_name":"AI-based Problem Solving and Planning","score":0.24199999868869781,"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/T10906","display_name":"AI-based Problem Solving and Planning","score":0.24199999868869781,"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/T11596","display_name":"Constraint Satisfaction and Optimization","score":0.23749999701976776,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"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/T11010","display_name":"Logic, Reasoning, and Knowledge","score":0.06610000133514404,"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/set","display_name":"Set (abstract data type)","score":0.2476000040769577},{"id":"https://openalex.org/keywords/noise","display_name":"Noise (video)","score":0.2443999946117401},{"id":"https://openalex.org/keywords/perception","display_name":"Perception","score":0.23739999532699585},{"id":"https://openalex.org/keywords/perspective","display_name":"Perspective (graphical)","score":0.23430000245571136}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.4650999903678894},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.38510000705718994},{"id":"https://openalex.org/C180747234","wikidata":"https://www.wikidata.org/wiki/Q23373","display_name":"Cognitive psychology","level":1,"score":0.37959998846054077},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.37450000643730164},{"id":"https://openalex.org/C77805123","wikidata":"https://www.wikidata.org/wiki/Q161272","display_name":"Social psychology","level":1,"score":0.26840001344680786},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.2476000040769577},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.2443999946117401},{"id":"https://openalex.org/C26760741","wikidata":"https://www.wikidata.org/wiki/Q160402","display_name":"Perception","level":2,"score":0.23739999532699585},{"id":"https://openalex.org/C12713177","wikidata":"https://www.wikidata.org/wiki/Q1900281","display_name":"Perspective (graphical)","level":2,"score":0.23430000245571136},{"id":"https://openalex.org/C2776325391","wikidata":"https://www.wikidata.org/wiki/Q6917865","display_name":"Motivated reasoning","level":3,"score":0.23399999737739563}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.acl-long.990","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.990","pdf_url":"https://aclanthology.org/2026.acl-long.990.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":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.acl-long.990","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.990","pdf_url":"https://aclanthology.org/2026.acl-long.990.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":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[{"id":"https://openalex.org/G4991685853","display_name":null,"funder_award_id":"62477012","funder_id":"https://openalex.org/F4320321001","funder_display_name":"National Natural Science Foundation of China"}],"funders":[{"id":"https://openalex.org/F4320321001","display_name":"National Natural Science Foundation of China","ror":"https://ror.org/01h0zpd94"},{"id":"https://openalex.org/F4320321133","display_name":"Chinese Academy of Sciences","ror":"https://ror.org/034t30j35"},{"id":"https://openalex.org/F4320321940","display_name":"Harbin Institute of Technology","ror":"https://ror.org/01yqg2h08"},{"id":"https://openalex.org/F4320325413","display_name":"Shanghai Municipal Commission of Economy and Informatization","ror":"https://ror.org/04166ws88"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166827583.pdf","grobid_xml":"https://content.openalex.org/works/W7166827583.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Large":[0],"reasoning":[1],"models":[2],"that":[3,42,74,90,107],"use":[4],"long":[5],"chainof-thought":[6],"excel":[7],"at":[8],"problem-solving":[9],"yet":[10],"waste":[11],"compute":[12],"on":[13,95,103],"redundant":[14],"checks.Curbing":[15],"this":[16,33],"overthinking":[17],"is":[18],"hard:":[19],"training-time":[20],"length":[21],"penalties":[22],"can":[23],"cripple":[24],"ability,":[25],"while":[26,127],"inferencetime":[27],"early-exit":[28,45],"adds":[29],"system":[30],"overhead.To":[31],"bridge":[32],"gap,":[34],"we":[35],"propose":[36],"Step-GRPO,":[37],"a":[38,69,86,110],"novel":[39],"post-training":[40],"framework":[41],"internalizes":[43],"dynamic":[44],"capabilities":[46],"directly":[47],"into":[48],"the":[49,52,76,124,129],"model.Step-GRPO":[50],"shifts":[51],"optimization":[53],"objective":[54],"from":[55],"raw":[56],"tokens":[57],"to":[58,65,78,123],"semantic":[59],"steps":[60],"by":[61,120],"utilizing":[62],"linguistic":[63],"markers":[64],"structure":[66],"reasoning.We":[67],"introduce":[68],"Dynamic":[70],"Truncated":[71],"Rollout":[72],"mechanism":[73],"exposes":[75],"model":[77,101,126],"concise":[79],"high-confidence":[80],"trajectories":[81],"during":[82],"exploration,":[83],"synergized":[84],"with":[85],"Step-Aware":[87],"Relative":[88],"Reward":[89],"dynamically":[91],"penalizes":[92],"redundancy":[93],"based":[94],"group-level":[96],"baselines.Extensive":[97],"experiments":[98],"across":[99],"three":[100],"sizes":[102],"diverse":[104],"benchmarks":[105],"demonstrate":[106],"Step-GRPO":[108],"achieves":[109],"superior":[111],"accuracy-efficiency":[112],"tradeoff.On":[113],"Qwen3-8B,":[114],"our":[115],"method":[116],"reduces":[117],"token":[118],"consumption":[119],"32.0%":[121],"compared":[122],"vanilla":[125],"avoiding":[128],"accuracy":[130],"degradation":[131],"observed":[132],"in":[133],"traditional":[134],"length-penalty":[135],"methods.":[136]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2026-07-02T00:00:00"}