[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-82589-en":3,"doc-seo-82589-105":30,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":21,"is_downloadable":21,"audit_status":21,"page_count":22,"language":23,"language_code":24,"site_id":25,"html_lang":24,"table_of_contents":26,"faqs":27,"seo_title":13,"seo_description":14,"update_tm":28,"read_time":29},82589,34359740700684,"Finn","https://ap-avatar.wpscdn.com/avatar/1f400023980c374ae676?_k=1777273430885731487",8,"Research & Report","Diffusion-GR2 Diffusion Generative Reasoning Re-ranker","Diffusion-GR2 converts an autoregressive generative reasoning re-ranker (GR2) into a faster block-diffusion re-ranker without relying on constrained decoders. Naive conversion creates structural errors (invalid permutations from parallel denoising) and distributional mismatch from off-policy fine-tuning. The method uses conversion fine-tuning to produce valid permutations, on-policy distillation to match the model’s own decoding trajectories using dense token targets, and a reinforcement-learning stage for reward-based re-ranking.","arXiv :2607 .0 1 170v4 [ cs .IR] 12 Jul 2026  \nDiffusion-GR2: Diffusion Generative Reasoning Re-ranker  \nZhuoxuan Zhang2 ,∗ , Kangqi Ni2 ,∗ , Yuhang Chen2 ,∗ , Mingfu Liang 1 ,∗ ,†,‡, Xiaohan Wei 1 , Yunchen Pu 1 , Fei Tian 1 , Chonglin Sun 1 , Frank Shyu 1 , Adam (Yang) Song 1 , Sandeep Pandey 1 , Luke Simon 1 ,‡, Tianlong Chen2 ,‡, Xi Liu 1 ,‡  \n1 Meta AI, 2 UNC Chapel Hill  \n∗ Co-first author, †Project and Execution Lead, ‡Joint corresponding author  \nGenerative reasoning re-rankers achieve strong recommendation accuracy by emitting a chain-ofthought before re-ordering a candidate list, but they are slow at inference: an autoregressive (AR) decoder spends one sequential forward pass per reasoning token, and the reasoning trace far exceeds the ranking it produces. To reduce this cost, block-diffusion language models decode many positions in parallel over a few denoising steps and are substantially faster, yet naively converting an AR re-ranker into one opens two accuracy gaps: (1) a structural gap: answer positions are denoised in parallel and scored independently, so the decoder emits invalid rankings (duplicated, dropped, or out-of-set identifiers) that AR avoids through left-to-right masking; and (2) a distributional gap: fine-tuning the converted model on fixed teacher trajectories is off-policy relative to its own decoding at inference, leaving a residual accuracy gap. To close both gaps while keeping the speedup, we propose Diffusion-GR2, a recipe that converts our AR reasoning re-ranker (GR2) into a block-diffusion re-ranker. First, conversion fine-tuning (CFT) adapts the AR-initialized diffusion model to denoise the answer into a valid permutation on its own, without an external constrained decoder. Next, on-policy distillation (OPD) then supervises the model on its own decoded trajectories with dense per-token targets from the AR teacher. Finally, we apply a reinforcement-learning (RL) stage against a re-ranking reward on top of OPD’s on-policy policy. Experiments on Amazon Beauty demonstrate that Diffusion-GR2 recovers to near-parity with the AR re-ranker, while block-parallel decoding raises decode throughput by 2.4–3.5 × at the model’s reasoning output length. Ablations show that CFT recovers most of the conversion gap, and that on-policy distillation further closes it to the AR reference.  \nDate: June 30, 2026  \nCorrespondence: Mingfu Liang ([mingfuliang@meta.com](mingfuliang@meta.com)), Luke Simon ([lukesimon@meta.com](lukesimon@meta.com)), Tianlong Chen  \n([tianlong@cs.unc.edu](tianlong@cs.unc.edu)), Xi Liu ([xliu1@meta.com](xliu1@meta.com))   \n1 Introduction  \nGenerative reasoning re-rankers (Liang et al. , 2026) built on large language models (LLMs) (Radford and Narasimhan, 2018 ; Radford et al. , 2019) have recently been shown to refine the final stage of recommendation pipelines by reasoning explicitly over a short candidate list before committing to an ordering. A reasoning re-ranker takes a user’s purchase history together with a pre-ranked list of candidate items produced by a retriever, generates a chain-of-thought that justifies a re-ordering, and emits a permutation of the candidates. The reasoning is what makes these models accurate: it grounds the decision in item semantics and user intent rather than a single learned score. It is also what makes them expensive. An autoregressive (AR) decoder spends one sequential forward pass per reasoning token, and the reasoning trace is far longer than the ranking it ultimately produces. In a production re-ranking setting, where the same model is queried for every impression, this sequential cost is the binding constraint.  \nBlock-diffusion language models (Arriola et al. , 2025b ; Wu et al. , 2025) suggest a way out. Rather than  \nDiffusion-GR2 Overview  \nConvert an accurate sequential GR2 re-ranker into a block-parallel diffusion reasoner  \nFigure 1 Overview of Diffusion-GR2. We start from the AR GR2 reasoning re-ranker, initialize a block-diffusion decoder f","cbCainJURrgGDsBR","https://ap.wps.com/l/cbCainJURrgGDsBR","pdf",944388,2,1,17,"English","en",105,"# Introduction\n## Motivation and Problem Setting\n## Proposed Method Overview\n### Conversion Fine-Tuning (CFT)\n### On-Policy Distillation (OPD)\n### Reinforcement Learning (RL)","[{\"question\":\"Why are generative reasoning re-rankers slow in inference?\",\"answer\":\"They generate a reasoning chain before producing a candidate ordering, and an autoregressive decoder runs one sequential forward pass per reasoning token. The reasoning trace is much longer than the final ranking.\"},{\"question\":\"What accuracy gaps occur when converting an AR re-ranker to a block-diffusion re-ranker naively?\",\"answer\":\"Two gaps arise: a structural gap where parallel denoising can yield invalid rankings (duplicated, dropped, or out-of-set identifiers), and a distributional gap because fixed teacher trajectories make fine-tuning off-policy relative to inference decoding.\"},{\"question\":\"How does Diffusion-GR2 recover accuracy while keeping speed?\",\"answer\":\"It applies conversion fine-tuning to denoise into valid permutations, then uses on-policy distillation to supervise on the model’s own decoded trajectories, and finally adds a reinforcement-learning stage guided by a re-ranking reward.\"}]",1784181676,43,{"code":4,"msg":31,"data":32},"ok",{"site_id":25,"language":24,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"diffusion-gr2-diffusion-generative-reasoning-re-ranker","",{"@graph":36,"@context":85},[37,53,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,47,50],{"item":41,"name":42,"@type":43,"position":21},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":20},"https://docshare.wps.com/document/","Document",{"item":48,"name":12,"@type":43,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":43,"position":52},"https://docshare.wps.com/document/diffusion-gr2-diffusion-generative-reasoning-re-ranker/82589/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":24,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":41,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-22","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"Why are generative reasoning re-rankers slow in inference?","Question",{"text":75,"@type":76},"They generate a reasoning chain before producing a candidate ordering, and an autoregressive decoder runs one sequential forward pass per reasoning token. The reasoning trace is much longer than the final ranking.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"What accuracy gaps occur when converting an AR re-ranker to a block-diffusion re-ranker naively?",{"text":80,"@type":76},"Two gaps arise: a structural gap where parallel denoising can yield invalid rankings (duplicated, dropped, or out-of-set identifiers), and a distributional gap because fixed teacher trajectories make fine-tuning off-policy relative to inference decoding.",{"name":82,"@type":73,"acceptedAnswer":83},"How does Diffusion-GR2 recover accuracy while keeping speed?",{"text":84,"@type":76},"It applies conversion fine-tuning to denoise into valid permutations, then uses on-policy distillation to supervise on the model’s own decoded trajectories, and finally adds a reinforcement-learning stage guided by a re-ranking reward.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":25},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":21,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":20,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":46,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":46,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":46,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]