{"id":"https://openalex.org/W7214799676","doi":"https://doi.org/10.48550/arxiv.2609.32457","title":"Write Back the $\u0394$: Revisiting the Same Tokens with Fresh Representations","display_name":"Write Back the $\u0394$: Revisiting the Same Tokens with Fresh Representations","publication_year":2026,"publication_date":"2026-09-26","ids":{"openalex":"https://openalex.org/W7214799676","doi":"https://doi.org/10.48550/arxiv.2609.32457"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2609.32457","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.32457","pdf_url":"https://arxiv.org/pdf/2609.32457","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2609.32457","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5123124139","display_name":"Wencheng Ye","orcid":"https://orcid.org/0009-0006-1718-6157","observed_orcids":["https://orcid.org/0009-0006-1718-6157"]},"institutions":[{"id":"https://openalex.org/I116953780","display_name":"Tongji University","ror":"https://ror.org/03rc6as71","country_code":"CN","type":"education","lineage":["https://openalex.org/I116953780"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Ye, Wencheng","raw_affiliation_strings":["Tongji University ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Tongji University ,","institution_ids":["https://openalex.org/I116953780"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5047787255","display_name":"Anning Hu","orcid":"https://orcid.org/0000-0002-3218-4954","observed_orcids":["https://orcid.org/0000-0002-3218-4954"]},"institutions":[{"id":"https://openalex.org/I183067930","display_name":"Shanghai Jiao Tong University","ror":"https://ror.org/0220qvk04","country_code":"CN","type":"education","lineage":["https://openalex.org/I183067930"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Hu, Anning","raw_affiliation_strings":["Shanghai Jiao Tong University ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Shanghai Jiao Tong University ,","institution_ids":["https://openalex.org/I183067930"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5153424014","display_name":"Xiangdong Zhang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I183067930","display_name":"Shanghai Jiao Tong University","ror":"https://ror.org/0220qvk04","country_code":"CN","type":"education","lineage":["https://openalex.org/I183067930"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Zhang, Xiangdong","raw_affiliation_strings":["Shanghai Jiao Tong University ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Shanghai Jiao Tong University ,","institution_ids":["https://openalex.org/I183067930"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5153402606","display_name":"Tianyi Wang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I116953780","display_name":"Tongji University","ror":"https://ror.org/03rc6as71","country_code":"CN","type":"education","lineage":["https://openalex.org/I116953780"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Wang, Tianyi","raw_affiliation_strings":["Tongji University ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Tongji University ,","institution_ids":["https://openalex.org/I116953780"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5153432953","display_name":"Yikang Li","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I4405279436","display_name":"Shanghai Innovation Institute","ror":"https://ror.org/01fatbw66","country_code":"CN","type":"education","lineage":["https://openalex.org/I4405279436"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Li, Yikang","raw_affiliation_strings":["Shanghai Innovation Institute"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Shanghai Innovation Institute","institution_ids":["https://openalex.org/I4405279436"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139834670","display_name":"Hengyu Jin","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I116953780","display_name":"Tongji University","ror":"https://ror.org/03rc6as71","country_code":"CN","type":"education","lineage":["https://openalex.org/I116953780"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Jin, Hengyu","raw_affiliation_strings":["Tongji University ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Tongji University ,","institution_ids":["https://openalex.org/I116953780"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5153520998","display_name":"Bing Li","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I116953780","display_name":"Tongji University","ror":"https://ror.org/03rc6as71","country_code":"CN","type":"education","lineage":["https://openalex.org/I116953780"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Li, Bing","raw_affiliation_strings":["Tongji University ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Tongji University ,","institution_ids":["https://openalex.org/I116953780"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5153445677","display_name":"Junchi Yan","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I183067930","display_name":"Shanghai Jiao Tong University","ror":"https://ror.org/0220qvk04","country_code":"CN","type":"education","lineage":["https://openalex.org/I183067930"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Yan, Junchi","raw_affiliation_strings":["Shanghai Jiao Tong University ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Shanghai Jiao Tong University ,","institution_ids":["https://openalex.org/I183067930"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":3,"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/T10028","display_name":"Topic Modeling","score":0.9082000255584717,"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":0.9082000255584717,"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":0.058400001376867294,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.006300000008195639,"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/residual-stream","display_name":"residual stream","score":0.4620000123977661},{"id":"https://openalex.org/keywords/reflux","display_name":"reflux","score":0.718999981880188},{"id":"https://openalex.org/keywords/transformer","display_name":"transformer models","score":0.33500000834465027},{"id":"https://openalex.org/keywords/multi-hop-reasoning","display_name":"multi-hop reasoning","score":0.7720000147819519},{"id":"https://openalex.org/keywords/language-modeling","display_name":"language modeling","score":0.7059999704360962}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7687000036239624},{"id":"https://openalex.org/C48044578","wikidata":"https://www.wikidata.org/wiki/Q727490","display_name":"Scalability","level":2,"score":0.7157999873161316},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.6747999787330627},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.6349999904632568},{"id":"https://openalex.org/C100279451","wikidata":"https://www.wikidata.org/wiki/Q372193","display_name":"Perplexity","level":3,"score":0.5641000270843506},{"id":"https://openalex.org/C155512373","wikidata":"https://www.wikidata.org/wiki/Q287450","display_name":"Residual","level":2,"score":0.5515000224113464},{"id":"https://openalex.org/C188198153","wikidata":"https://www.wikidata.org/wiki/Q1613840","display_name":"Limiting","level":2,"score":0.460999995470047},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.3882000148296356},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.38359999656677246},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.3285999894142151},{"id":"https://openalex.org/C132525143","wikidata":"https://www.wikidata.org/wiki/Q141488","display_name":"Graph","level":2,"score":0.3230000138282776},{"id":"https://openalex.org/C186886427","wikidata":"https://www.wikidata.org/wiki/Q5441213","display_name":"Feedback loop","level":2,"score":0.31360000371932983},{"id":"https://openalex.org/C42058472","wikidata":"https://www.wikidata.org/wiki/Q810214","display_name":"Base (topology)","level":2,"score":0.31209999322891235},{"id":"https://openalex.org/C38858127","wikidata":"https://www.wikidata.org/wiki/Q5441228","display_name":"Feed forward","level":2,"score":0.3111000061035156},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.30390000343322754},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.288100004196167},{"id":"https://openalex.org/C2779843651","wikidata":"https://www.wikidata.org/wiki/Q7390335","display_name":"SIGNAL (programming language)","level":2,"score":0.2687999904155731},{"id":"https://openalex.org/C146380142","wikidata":"https://www.wikidata.org/wiki/Q1137726","display_name":"Directed graph","level":2,"score":0.2540999948978424},{"id":"https://openalex.org/C68649174","wikidata":"https://www.wikidata.org/wiki/Q1379116","display_name":"Base station","level":2,"score":0.2515999972820282},{"id":"https://openalex.org/C120314980","wikidata":"https://www.wikidata.org/wiki/Q180634","display_name":"Distributed computing","level":1,"score":0.25110000371932983}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2609.32457","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.32457","pdf_url":"https://arxiv.org/pdf/2609.32457","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2609.32457","doi":"https://doi.org/10.48550/arxiv.2609.32457","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.32457","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2609.32457","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.32457","pdf_url":"https://arxiv.org/pdf/2609.32457","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7214799676.pdf","grobid_xml":"https://content.openalex.org/works/W7214799676.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7213918674","https://openalex.org/W7163582471","https://openalex.org/W7202362188","https://openalex.org/W7214896883","https://openalex.org/W7214290239","https://openalex.org/W7170442493","https://openalex.org/W7140222370","https://openalex.org/W4415020698","https://openalex.org/W4415272785","https://openalex.org/W7165397770"],"abstract_inverted_index":{"Transformers":[0],"process":[1],"information":[2],"strictly":[3],"forward":[4,20],"through":[5],"depth,":[6,26],"preventing":[7],"deeper":[8,56],"computation":[9,53,81],"from":[10],"revisiting":[11],"and":[12,90,112,124,136,149],"refining":[13],"earlier":[14,61],"representations.":[15,205],"To":[16],"augment":[17],"the":[18,32,74,95,121,173,190,200],"standard":[19],"pass,":[21],"existing":[22],"approaches":[23],"either":[24],"re-execute":[25],"incurring":[27],"additional":[28],"computation,":[29],"or":[30],"modify":[31],"residual":[33,57],"stream":[34],"using":[35],"predefined":[36],"directions,":[37],"limiting":[38],"their":[39],"instance-level":[40],"adaptation.":[41],"Recently,":[42],"inference-time":[43],"feedback":[44,92,107,119,126],"offers":[45],"a":[46,86,105],"direct":[47],"mechanism":[48],"for":[49,188],"recycling":[50],"endogenously":[51],"produced":[52],"by":[54,152],"writing":[55],"states":[58],"back":[59,68],"to":[60,120,127,198],"layers,":[62,84],"yet":[63],"what":[64],"should":[65],"be":[66,209],"fed":[67],"remains":[69],"unclear.":[70],"We":[71],"argue":[72],"that":[73,109,139],"depth":[75],"increment":[76],"Delta,":[77],"capturing":[78],"newly":[79],"accumulated":[80],"between":[82],"two":[83],"provides":[85],"more":[87],"effective,":[88],"composable,":[89],"scalable":[91],"signal":[93],"than":[94],"full":[96],"state.":[97],"Building":[98],"on":[99,160],"this":[100],"observation,":[101],"we":[102],"introduce":[103],"ReFlux,":[104],"learnable":[106],"graph":[108],"dynamically":[110],"selects":[111],"composes":[113],"increment-carrying":[114],"routes.":[115],"ReFlux":[116,141,164,183],"supports":[117],"synchronous":[118,140],"same":[122,201],"token":[123],"streaming":[125],"subsequent":[128],"tokens.":[129],"Extensive":[130],"experiments":[131],"across":[132,145],"various":[133],"models,":[134],"corpora,":[135,148],"benchmarks":[137],"show":[138],"consistently":[142],"reduces":[143],"perplexity":[144],"ten":[146],"language-modeling":[147],"improves":[150],"accuracy":[151],"2.1-2.3":[153],"points,":[154],"with":[155,203],"gains":[156,170],"reaching":[157],"4.7":[158],"points":[159],"multi-hop":[161],"reasoning.":[162],"Streaming":[163],"further":[165],"retains":[166],"most":[167],"of":[168,194],"these":[169],"while":[171],"preserving":[172],"base":[174],"model's":[175],"1x":[176],"theoretical":[177],"backbone":[178],"FLOPs.":[179],"These":[180],"results":[181],"establish":[182],"as":[184],"an":[185],"efficient":[186],"paradigm":[187],"unlocking":[189],"latent":[191],"computational":[192],"potential":[193],"LLMs,":[195],"allowing":[196],"them":[197],"revisit":[199],"tokens":[202],"fresh":[204],"Code":[206],"implementation":[207],"can":[208],"found":[210],"at":[211],"https://github.com/gooogleshanghai/reflux.":[212]},"counts_by_year":[],"updated_date":"2026-10-28T10:33:24.094632","created_date":"2026-09-30T00:00:00"}