[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-327248-105":3,"detail-sidebar-cat-0-en-105":73,"doc-detail-327248-en":123},{"code":4,"msg":5,"data":6},0,"ok",{"site_id":7,"language":8,"slug":9,"title":10,"keywords":11,"description":12,"schema_data":13,"social_meta":66,"head_meta":68,"extra_data":70,"updated_unix":72},105,"en","diffusion-gr2-diffusion-generative-reasoning-re-ranker-327248","Diffusion-GR2 - Diffusion Generative Reasoning Re-ranker","","Diffusion-GR2 converts an autoregressive generative-reasoning re-ranker into a block-diffusion re-ranker to improve inference speed without sacrificing ranking quality. The method targets two accuracy losses from conversion: a structural gap where parallel denoising can yield invalid permutations, and a distributional gap from off-policy fine-tuning on fixed teacher trajectories. Conversion fine-tuning restores valid permutation denoising, on-policy distillation matches self-generated trajectories using dense token targets from the AR teacher, and an RL stage optimizes the re-ranking reward. Experiments on Amazon Beauty recover near parity while achieving 2.4–3.5× decode throughput.",{"@graph":14,"@context":65},[15,34,56],{"@type":16,"itemListElement":17},"BreadcrumbList",[18,23,27,31],{"item":19,"name":20,"@type":21,"position":22},"https://docshare.wps.com","Home","ListItem",1,{"item":24,"name":25,"@type":21,"position":26},"https://docshare.wps.com/document/","Document",2,{"item":28,"name":29,"@type":21,"position":30},"https://docshare.wps.com/document/research-report/","Research & Report",3,{"item":32,"name":10,"@type":21,"position":33},"https://docshare.wps.com/document/diffusion-gr2-diffusion-generative-reasoning-re-ranker-327248/327248/",4,{"url":32,"name":10,"@type":35,"image":36,"author":41,"headline":10,"publisher":44,"fileFormat":47,"inLanguage":8,"description":12,"dateModified":48,"datePublished":49,"encodingFormat":47,"isAccessibleForFree":50,"interactionStatistic":51},"DigitalDocument",{"url":37,"@type":38,"width":39,"height":40},"https://docshare.wps.com/thumbnails/diffusion-gr2-diffusion-generative-reasoning-re-ranker-327248/327248.png","ImageObject",300,407,{"name":42,"@type":43},"Lucas Martin","Person",{"url":19,"name":45,"@type":46},"DocShare","Organization","application/pdf","2026-10-27","2026-09-21",true,{"@type":52,"interactionType":53,"userInteractionCount":55},"InteractionCounter",{"@type":54},"ViewAction",13,{"@type":57,"mainEntity":58},"FAQPage",[59],{"name":60,"@type":61,"acceptedAnswer":62},"How does Diffusion-GR2 close the structural and distributional gaps?","Question",{"text":63,"@type":64},"Conversion fine-tuning adapts the diffusion model so denoising produces a valid permutation without an external constrained decoder. On-policy distillation then supervises the model on its own decoded trajectories with dense per-token targets from the AR teacher, and a subsequent RL stage optimizes a re-ranking reward.","Answer","https://schema.org",{"og:url":32,"og:type":67,"og:title":10,"og:site_name":45,"og:description":12},"article",{"robots":69,"canonical":32},"index,follow",{"doc_id":71,"site_id":7},327248,1790218648,{"code":4,"msg":74,"data":75},"success",[76,80,84,88,93,98,103,107,112,115,119],{"id":22,"doc_module":4,"doc_module_name":25,"category_name":77,"show_sort_weight":78,"slug":79},"Story & Novel",90,"story-novel",{"id":26,"doc_module":4,"doc_module_name":25,"category_name":81,"show_sort_weight":82,"slug":83},"Literature",80,"literature",{"id":33,"doc_module":4,"doc_module_name":25,"category_name":85,"show_sort_weight":86,"slug":87},"Exam",70,"exam",{"id":89,"doc_module":4,"doc_module_name":25,"category_name":90,"show_sort_weight":91,"slug":92},5,"Comic",60,"comic",{"id":94,"doc_module":4,"doc_module_name":25,"category_name":95,"show_sort_weight":96,"slug":97},6,"Technology",50,"technology",{"id":99,"doc_module":4,"doc_module_name":25,"category_name":100,"show_sort_weight":101,"slug":102},7,"Healthcare",40,"healthcare",{"id":104,"doc_module":4,"doc_module_name":25,"category_name":29,"show_sort_weight":105,"slug":106},8,30,"research-report",{"id":108,"doc_module":4,"doc_module_name":25,"category_name":109,"show_sort_weight":110,"slug":111},9,"Religion & Spirituality",20,"religion-spirituality",{"id":110,"doc_module":4,"doc_module_name":25,"category_name":113,"show_sort_weight":110,"slug":114},"World Cup","world-cup",{"id":116,"doc_module":4,"doc_module_name":25,"category_name":117,"show_sort_weight":116,"slug":118},10,"Lifestyle","lifestyle",{"id":120,"doc_module":4,"doc_module_name":25,"category_name":121,"show_sort_weight":89,"slug":122},19,"General","general",{"code":4,"msg":74,"data":124},{"doc_id":71,"user_id":125,"nickname":42,"user_avatar":126,"doc_module":4,"category_id":104,"category_name":29,"doc_title":10,"doc_description":12,"doc_content":127,"file_id":128,"file_url":129,"file_type":130,"file_size":131,"view_count":55,"is_deleted":4,"is_public":22,"is_downloadable":22,"audit_status":22,"page_count":132,"language":133,"language_code":8,"site_id":7,"html_lang":8,"table_of_contents":134,"faqs":135,"seo_title":136,"seo_description":12,"update_tm":137,"read_time":138,"preview_page_count":4},8796095360427,"https://ap-avatar.wpscdn.com/davatar_994ba38a5ba835b3df7d355c54d3ed8d","arXiv :2607 .0 1 170v2 [ cs .IR] 4 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  \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 walking the sequence left to right, they decode many positions at once and refine them over a small number of denoising steps, committing the most confident positions at each step. This is attractive precisely where ARis most expensive, i.e., the long reasoning span, and it is the source of a large potential speedup. But the  \nDiffusion-GR2 Overview  \nConvert an accurate sequential GR2 re-ranker into a block-parallel diffusion reasoner  \nFigure 1 Overview of Diffusion-GR","cbCairtmuQWKbG1G","https://ap.wps.com/l/cbCairtmuQWKbG1G","pdf",1032924,17,"English","# Introduction\n## Overview of Diffusion-GR2\n## Conversion, distillation, and reinforcement learning stages","[{\"question\":\"How does Diffusion-GR2 close the structural and distributional gaps?\",\"answer\":\"Conversion fine-tuning adapts the diffusion model so denoising produces a valid permutation without an external constrained decoder. On-policy distillation then supervises the model on its own decoded trajectories with dense per-token targets from the AR teacher, and a subsequent RL stage optimizes a re-ranking reward.\"}]","Diffusion-GR2 - Diffusion Generative Reasoning Re-ranker | PDF",1789976806,43]