{"id":"https://openalex.org/W7223701375","doi":"https://doi.org/10.66977/xsci.2610.002r","title":"On-Policy Residual Learning for Diffusion Language Models","display_name":"On-Policy Residual Learning for Diffusion Language Models","publication_year":2026,"publication_date":"2026-10-26","ids":{"openalex":"https://openalex.org/W7223701375","doi":"https://doi.org/10.66977/xsci.2610.002r"},"language":"en","primary_location":{"id":"doi:10.66977/xsci.2610.002r","doi":"https://doi.org/10.66977/xsci.2610.002r","is_oa":true,"landing_page_url":"https://doi.org/10.66977/xsci.2610.002r","pdf_url":"https://xsci.com/preprint/doi/10.66977/xsci.2610.002r.pdf","source":null,"license":null,"license_id":null,"version":"acceptedVersion","is_accepted":true,"is_published":false,"raw_source_name":null,"raw_type":"posted-content"},"volume":null,"type":"preprint","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://xsci.com/preprint/doi/10.66977/xsci.2610.002r.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Anonymous","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Anonymous","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":null,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.5849999785423279,"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":0.5849999785423279,"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":0.3240000009536743,"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/T10775","display_name":"Generative Adversarial Networks and Image Synthesis","score":0.020099999383091927,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"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-language-models","display_name":"diffusion language models","score":0.9269999861717224},{"id":"https://openalex.org/keywords/speculative-decoding","display_name":"speculative decoding","score":0.6000000238418579},{"id":"https://openalex.org/keywords/llada","display_name":"LLaDA","score":0.621999979019165},{"id":"https://openalex.org/keywords/qwen3-4b","display_name":"Qwen3-4B","score":0.6510000228881836}],"concepts":[{"id":"https://openalex.org/C155512373","wikidata":"https://www.wikidata.org/wiki/Q287450","display_name":"Residual","level":2,"score":0.8697999715805054},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.6753000020980835},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6437000036239624},{"id":"https://openalex.org/C137635306","wikidata":"https://www.wikidata.org/wiki/Q182667","display_name":"Pareto principle","level":2,"score":0.5005000233650208},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.48730000853538513},{"id":"https://openalex.org/C2781137444","wikidata":"https://www.wikidata.org/wiki/Q237105","display_name":"Sharpening","level":2,"score":0.4406000077724457},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.42309999465942383},{"id":"https://openalex.org/C2781002164","wikidata":"https://www.wikidata.org/wiki/Q6822311","display_name":"Meta learning (computer science)","level":3,"score":0.39559999108314514},{"id":"https://openalex.org/C157764524","wikidata":"https://www.wikidata.org/wiki/Q1383412","display_name":"Throughput","level":3,"score":0.3921999931335449},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3880000114440918},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.3799999952316284},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.3499000072479248},{"id":"https://openalex.org/C111335779","wikidata":"https://www.wikidata.org/wiki/Q3454686","display_name":"Reduction (mathematics)","level":2,"score":0.3131999969482422},{"id":"https://openalex.org/C2164484","wikidata":"https://www.wikidata.org/wiki/Q5170150","display_name":"Core (optical fiber)","level":2,"score":0.31200000643730164},{"id":"https://openalex.org/C151319957","wikidata":"https://www.wikidata.org/wiki/Q752739","display_name":"Asynchronous communication","level":2,"score":0.31119999289512634},{"id":"https://openalex.org/C774472","wikidata":"https://www.wikidata.org/wiki/Q6760393","display_name":"Margin (machine learning)","level":2,"score":0.30480000376701355},{"id":"https://openalex.org/C112972136","wikidata":"https://www.wikidata.org/wiki/Q7595718","display_name":"Stability (learning theory)","level":2,"score":0.2806999981403351},{"id":"https://openalex.org/C185798385","wikidata":"https://www.wikidata.org/wiki/Q1161707","display_name":"Benchmark (surveying)","level":2,"score":0.2655999958515167},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.2547999918460846},{"id":"https://openalex.org/C61797465","wikidata":"https://www.wikidata.org/wiki/Q1188986","display_name":"Term (time)","level":2,"score":0.25209999084472656}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.66977/xsci.2610.002r","doi":"https://doi.org/10.66977/xsci.2610.002r","is_oa":true,"landing_page_url":"https://doi.org/10.66977/xsci.2610.002r","pdf_url":"https://xsci.com/preprint/doi/10.66977/xsci.2610.002r.pdf","source":null,"license":null,"license_id":null,"version":"acceptedVersion","is_accepted":true,"is_published":false,"raw_source_name":null,"raw_type":"posted-content"}],"best_oa_location":{"id":"doi:10.66977/xsci.2610.002r","doi":"https://doi.org/10.66977/xsci.2610.002r","is_oa":true,"landing_page_url":"https://doi.org/10.66977/xsci.2610.002r","pdf_url":"https://xsci.com/preprint/doi/10.66977/xsci.2610.002r.pdf","source":null,"license":null,"license_id":null,"version":"acceptedVersion","is_accepted":true,"is_published":false,"raw_source_name":null,"raw_type":"posted-content"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7223701375.pdf","grobid_xml":"https://content.openalex.org/works/W7223701375.grobid-xml"},"referenced_works_count":41,"referenced_works":["https://openalex.org/W2866343820","https://openalex.org/W3134642945","https://openalex.org/W3177813494","https://openalex.org/W3195140297","https://openalex.org/W3210904070","https://openalex.org/W4300425011","https://openalex.org/W4312122975","https://openalex.org/W4388718089","https://openalex.org/W4392822190","https://openalex.org/W4403883066","https://openalex.org/W4405903187","https://openalex.org/W4407679583","https://openalex.org/W4414856864","https://openalex.org/W4414939355","https://openalex.org/W4415097124","https://openalex.org/W4416048624","https://openalex.org/W4416048648","https://openalex.org/W4416051103","https://openalex.org/W6910504028","https://openalex.org/W7092291868","https://openalex.org/W7117584765","https://openalex.org/W7123762124","https://openalex.org/W7127130698","https://openalex.org/W7128006584","https://openalex.org/W7128507389","https://openalex.org/W7128793353","https://openalex.org/W7137924825","https://openalex.org/W7152996795","https://openalex.org/W7154266495","https://openalex.org/W7162290526","https://openalex.org/W7162430223","https://openalex.org/W7162773333","https://openalex.org/W7164848801","https://openalex.org/W7165002948","https://openalex.org/W7167577630","https://openalex.org/W7167612858","https://openalex.org/W7167729652","https://openalex.org/W7172402159","https://openalex.org/W7172536370","https://openalex.org/W7196954301","https://openalex.org/W7203646259"],"related_works":["https://openalex.org/W7222576614","https://openalex.org/W7161832888","https://openalex.org/W7202412525","https://openalex.org/W7127130698","https://openalex.org/W7163900200","https://openalex.org/W7220719280","https://openalex.org/W7140223010","https://openalex.org/W7118527898","https://openalex.org/W7125531142","https://openalex.org/W7161172251"],"abstract_inverted_index":{"Residual":[0],"context":[1,60],"(DiffusionGemma":[2],"Team":[3],"et":[4],"al.,":[5],"2026)":[6],"is":[7,48,53,81],"emerging":[8],"as":[9,104],"a":[10,67,152,193,275,295],"key":[11],"ingredient":[12],"for":[13,57,69],"unlocking":[14],"the":[15,93,105,128],"speed\u2013quality":[16,195],"Pareto":[17,196],"frontier":[18,197],"of":[19,38],"diffusion":[20],"language":[21],"models":[22,200,273],"(DLMs)":[23],"by":[24,164,176,265],"carrying":[25],"forward":[26,162],"predictive":[27],"distributions":[28],"across":[29],"denoising":[30],"steps.":[31],"Yet":[32],"turning":[33],"this":[34,190],"seemingly":[35],"simple":[36],"mechanism":[37],"computation":[39],"recirculation":[40],"into":[41],"an":[42,97],"effective":[43],"and":[44,95,119,135,140,204,211,232,250,259,269,291],"efficient":[45],"learning":[46,72,80,129,147],"paradigm":[47],"far":[49],"from":[50,87,156],"simple:":[51],"there":[52],"no":[54],"clear":[55],"recipe":[56],"how":[58],"residual":[59,71,79,90,114,133,146],"should":[61],"be":[62],"learned.":[63],"We":[64],"introduce":[65],"ORL,":[66],"framework":[68],"on-policy":[70,145],"in":[73],"DLMs.":[74],"Algorithmically,":[75],"we":[76],"identify":[77],"that":[78,100],"inherently":[82],"on-policy.":[83],"ORL":[84,126,158,222,262,272],"learns":[85],"directly":[86],"these":[88],"model-induced":[89],"trajectories":[91],"on":[92,187,228,260,283,288,294],"fly":[94],"introduces":[96],"entropy-contraction":[98],"objective":[99],"encourages":[101],"sharpening":[102],"predictions":[103],"model":[106,191],"revisits":[107],"its":[108],"own":[109],"intermediate":[110,124],"predictions.":[111],"System-wise,":[112],"exact":[113],"supervision":[115],"requires":[116],"capturing,":[117],"retaining,":[118],"repeatedly":[120],"processing":[121],"many":[122],"trajectory-dependent":[123],"states.":[125],"co-designs":[127],"algorithm":[130],"with":[131,169,202,299],"asynchronous":[132],"capture":[134],"transport,":[136],"specialized":[137],"block-diffusion":[138],"attention,":[139],"packed":[141],"compute-balanced":[142],"training,":[143],"making":[144],"practical":[148],"at":[149,179,219,256,278],"scale.":[150],"On":[151],"4B":[153],"DLM":[154],"converted":[155],"Qwen3-4B,":[157],"improves":[159],"tokens":[160],"per":[161],"(TPF)":[163],"up":[165,245,251,279],"to":[166,225,246,252,280],"6.2\u00d7":[167],"and,":[168],"our":[170],"proposed":[171],"DLMRT":[172,268],"engine,":[173],"end-to-end":[174],"throughput":[175,255],"approximately":[177],"6.9\u00d7":[178],"comparable":[180,257,300],"or":[181],"better":[182],"quality.":[183,301],"In":[184],"single-request":[185],"serving":[186],"one":[188,284],"H100,":[189],"achieves":[192],"stronger":[194],"than":[198,217,240],"autoregressive":[199],"accelerated":[201],"EAGLE3":[203],"DFlash":[205,218],"speculative":[206],"decoding":[207],"under":[208],"both":[209],"strict":[210],"relaxed":[212],"verification,":[213],"running":[214],"1.74\u00d7":[215],"faster":[216],"matched":[220],"accuracy.":[221],"also":[223],"transfers":[224],"existing":[226],"DLMs:":[227],"LLaDA-2.1-mini,":[229],"36":[230],"H100-hours":[231],"4.2M":[233],"self-generated":[234],"tokens,":[235],"about":[236],"10\u00d7":[237],"less":[238],"data":[239],"prior":[241],"LLaDA":[242],"recipes,":[243],"yield":[244],"1.5\u00d7":[247],"higher":[248,254],"TPF":[249,264],"3\u00d7":[253],"quality,":[258],"DiffusionGemma,":[261],"raises":[263],"1.1\u20131.2\u00d7.":[266],"With":[267],"NVFP4":[270],"quantization,":[271],"decode":[274],"single":[276],"request":[277],"5,046":[281],"tokens/s":[282,287,293],"B200,":[285],"7,708":[286],"eight":[289],"B200s,":[290],"2,788":[292],"consumer":[296],"RTX":[297],"5090,":[298]},"counts_by_year":[],"updated_date":"2026-11-04T10:54:05.112534","created_date":"2026-10-27T00:00:00"}