{"id":"https://openalex.org/W7166803497","doi":"https://doi.org/10.18653/v1/2026.acl-long.1723","title":"TA-GRPO-d: Trajectory-Aware GRPO for Optimizing Denoising Trajectories in Diffusion LLMs","display_name":"TA-GRPO-d: Trajectory-Aware GRPO for Optimizing Denoising Trajectories in Diffusion LLMs","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166803497","doi":"https://doi.org/10.18653/v1/2026.acl-long.1723"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2026.acl-long.1723","doi":"https://doi.org/10.18653/v1/2026.acl-long.1723","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1723","pdf_url":"https://aclanthology.org/2026.acl-long.1723.pdf","source":{"id":"https://openalex.org/S4393915740","display_name":"Proceedings of the conference - Association for Computational Linguistics. Meeting","issn_l":"0736-587X","issn":["0736-587X"],"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":["norway-1"],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"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"},"volume":null,"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"hybrid","oa_url":"https://aclanthology.org/2026.acl-long.1723.pdf","any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5021541115","display_name":"Gyunyeop Kim","orcid":"https://orcid.org/0000-0002-9604-8134","observed_orcids":["https://orcid.org/0000-0002-9604-8134"]},"institutions":[{"id":"https://openalex.org/I12832649","display_name":"Gachon University","ror":"https://ror.org/03ryywt80","country_code":"KR","type":"education","lineage":["https://openalex.org/I12832649"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Gyunyeop Kim","raw_affiliation_strings":["Department of Computing , Gachon University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Computing , Gachon University","institution_ids":["https://openalex.org/I12832649"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5059572713","display_name":"Sangwoo Kang","orcid":"https://orcid.org/0000-0002-0281-1726","observed_orcids":["https://orcid.org/0000-0002-0281-1726","https://orcid.org/0009-0005-4573-3210"]},"institutions":[{"id":"https://openalex.org/I12832649","display_name":"Gachon University","ror":"https://ror.org/03ryywt80","country_code":"KR","type":"education","lineage":["https://openalex.org/I12832649"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Sangwoo Kang","raw_affiliation_strings":["Department of Computing , Gachon University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Computing , Gachon University","institution_ids":["https://openalex.org/I12832649"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I12832649"],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.3971546,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"37159","last_page":"37174"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.5810999870300293,"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":0.5810999870300293,"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":0.32030001282691956,"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":0.019999999552965164,"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/diffusion-large-language-models","display_name":"diffusion large language models","score":0.949999988079071},{"id":"https://openalex.org/keywords/group-relative-policy-optimization","display_name":"Group Relative Policy Optimization","score":0.9129999876022339},{"id":"https://openalex.org/keywords/decoding-efficiency","display_name":"decoding efficiency","score":0.6349999904632568},{"id":"https://openalex.org/keywords/mathematical-reasoning","display_name":"mathematical reasoning","score":0.35899999737739563},{"id":"https://openalex.org/keywords/code-generation","display_name":"code generation","score":0.8259999752044678}],"concepts":[{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.4205000102519989},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.40310001373291016},{"id":"https://openalex.org/C69357855","wikidata":"https://www.wikidata.org/wiki/Q163214","display_name":"Diffusion","level":2,"score":0.3637000024318695},{"id":"https://openalex.org/C13662910","wikidata":"https://www.wikidata.org/wiki/Q193139","display_name":"Trajectory","level":2,"score":0.3598000109195709},{"id":"https://openalex.org/C28826006","wikidata":"https://www.wikidata.org/wiki/Q33521","display_name":"Applied mathematics","level":1,"score":0.3441999852657318},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.3425999879837036},{"id":"https://openalex.org/C163294075","wikidata":"https://www.wikidata.org/wiki/Q581861","display_name":"Noise reduction","level":2,"score":0.3409000039100647},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.31119999289512634},{"id":"https://openalex.org/C112972136","wikidata":"https://www.wikidata.org/wiki/Q7595718","display_name":"Stability (learning theory)","level":2,"score":0.2946000099182129},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.2736000120639801},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.2685000002384186},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.2612000107765198},{"id":"https://openalex.org/C111335779","wikidata":"https://www.wikidata.org/wiki/Q3454686","display_name":"Reduction (mathematics)","level":2,"score":0.2556999921798706},{"id":"https://openalex.org/C47446073","wikidata":"https://www.wikidata.org/wiki/Q5165890","display_name":"Control theory (sociology)","level":3,"score":0.2522999942302704},{"id":"https://openalex.org/C149782125","wikidata":"https://www.wikidata.org/wiki/Q160039","display_name":"Econometrics","level":1,"score":0.25110000371932983}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.acl-long.1723","doi":"https://doi.org/10.18653/v1/2026.acl-long.1723","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1723","pdf_url":"https://aclanthology.org/2026.acl-long.1723.pdf","source":{"id":"https://openalex.org/S4393915740","display_name":"Proceedings of the conference - Association for Computational Linguistics. Meeting","issn_l":"0736-587X","issn":["0736-587X"],"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":["norway-1"],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"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.1723","doi":"https://doi.org/10.18653/v1/2026.acl-long.1723","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1723","pdf_url":"https://aclanthology.org/2026.acl-long.1723.pdf","source":{"id":"https://openalex.org/S4393915740","display_name":"Proceedings of the conference - Association for Computational Linguistics. Meeting","issn_l":"0736-587X","issn":["0736-587X"],"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":["norway-1"],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"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":[],"study_designs":[],"awards":[{"id":"https://openalex.org/G8351729092","display_name":null,"funder_award_id":"2022R1A2C1005316","funder_id":"https://openalex.org/F4320328359","funder_display_name":"Ministry of Science and ICT, South Korea"},{"id":"https://openalex.org/G8801488284","display_name":null,"funder_award_id":"2022R1A2C1005316","funder_id":"https://openalex.org/F4320322120","funder_display_name":"National Research Foundation of Korea"}],"funders":[{"id":"https://openalex.org/F4320320671","display_name":"National Research Foundation","ror":"https://ror.org/05s0g1g46"},{"id":"https://openalex.org/F4320322120","display_name":"National Research Foundation of Korea","ror":"https://ror.org/013aysd81"},{"id":"https://openalex.org/F4320328359","display_name":"Ministry of Science and ICT, South Korea","ror":"https://ror.org/01wpjm123"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166803497.pdf","grobid_xml":"https://content.openalex.org/works/W7166803497.grobid-xml"},"referenced_works_count":17,"referenced_works":["https://openalex.org/W2736601468","https://openalex.org/W3181640983","https://openalex.org/W3195140297","https://openalex.org/W4226278401","https://openalex.org/W4308671448","https://openalex.org/W4322718191","https://openalex.org/W4327810158","https://openalex.org/W4379089709","https://openalex.org/W4407679583","https://openalex.org/W4414281281","https://openalex.org/W4415311663","https://openalex.org/W4415345686","https://openalex.org/W4416051103","https://openalex.org/W4416543926","https://openalex.org/W6904055037","https://openalex.org/W6948154002","https://openalex.org/W6966870376"],"related_works":["https://openalex.org/W7161741183","https://openalex.org/W4414971020","https://openalex.org/W7163024699","https://openalex.org/W7139914875","https://openalex.org/W7106243038","https://openalex.org/W4414825091","https://openalex.org/W7136150933","https://openalex.org/W7166846310","https://openalex.org/W7163538502","https://openalex.org/W7216523591"],"abstract_inverted_index":{"Diffusion":[0],"large":[1],"language":[2],"models":[3],"(dLLMs)":[4],"generate":[5],"text":[6],"by":[7,157],"repeatedly":[8],"unmasking":[9,41],"a":[10,45,55,60,93,99,111],"partially":[11],"noised":[12],"sequence":[13],"in":[14],"parallel,":[15],"promising":[16],"lower":[17,166],"latency":[18],"than":[19],"autoregressive":[20],"decoding.However,":[21],"most":[22],"discrete":[23],"dLLMs":[24],"still":[25],"rely":[26],"on":[27,84,136],"fixed":[28],"denoising":[29,62,82,155],"schedules,":[30],"which":[31,67,97],"are":[32],"non-adaptive":[33],"to":[34,69,74,119,159],"input":[35],"difficulty":[36],"and":[37,71,140,165],"cannot":[38],"learn":[39,120],"efficient":[40,181],"orders.This":[42],"paper":[43],"introduces":[44],"reinforcement":[46],"learning":[47],"(RL)":[48],"framework":[49,173],"that":[50,64,146],"transforms":[51],"dLLM":[52,176],"decoding":[53,133,177],"into":[54,92],"trajectory-aware,":[56],"learnable":[57],"policy.We":[58],"propose":[59],"confidence-gated":[61],"strategy":[63],"dynamically":[65],"decides":[66],"tokens":[68],"unmask":[70,75],"how":[72],"many":[73],"per":[76],"step,":[77],"enabling":[78],"adaptive":[79],"exploration":[80],"of":[81,105,131],"trajectories.Building":[83],"Group":[85],"Relative":[86],"Policy":[87],"Optimization,":[88],"we":[89],"reformulate":[90],"it":[91],"trajectory-aware":[94],"variant,":[95],"TA-GRPO-d,":[96],"combines":[98],"trajectory-level":[100],"signal-captured":[101],"as":[102],"the":[103,106,117,123,129,132],"z-score":[104],"AUC":[107],"over":[108],"intermediate":[109],"rewards-with":[110],"token-level":[112],"unmasking-time":[113],"weight.This":[114],"design":[115],"allows":[116],"model":[118],"not":[121],"only":[122],"final":[124],"output":[125],"quality":[126],"but":[127],"also":[128],"efficiency":[130],"path":[134],"itself.Experiments":[135],"MATH-500,":[137],"Countdown,":[138],"Sudoku,":[139],"code":[141],"benchmarks":[142],"(HumanEval,":[143],"MBPP)":[144],"show":[145],"TA-GRPO-d":[147],"maintains":[148],"or":[149],"improves":[150],"accuracy":[151],"while":[152],"reducing":[153],"average":[154],"steps":[156],"up":[158],"half,":[160],"achieving":[161],"both":[162],"faster":[163],"inference":[164],"computational":[167],"cost.Our":[168],"approach":[169],"provides":[170],"an":[171],"RL":[172],"for":[174],"optimizing":[175],"policies":[178],"toward":[179],"adaptive,":[180],"reasoning.":[182]},"counts_by_year":[],"updated_date":"2026-11-04T10:54:05.112534","created_date":"2026-07-02T00:00:00"}