[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-84799-en":3,"doc-seo-84799-105":29,"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":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},84799,5909877438554,"Maeve","https://ap-avatar.wpscdn.com/avatar/5600025385ad2bf12a7?_k=1778553567797529272",8,"Research & Report","Noisy Channel Minimum Bayes Risk Decoding","Noisy-Channel Minimum Bayes Risk (MBR) decoding improves text generation robustness and quality by maximizing expected utility over sampled pseudo-references rather than selecting the most probable hypothesis as in MAP decoding. The work addresses a key mismatch: original MBR conditions expected utility on pseudo-references, while common metrics like BLEU and COMET are asymmetric. It introduces a noisy-channel decomposition capturing bidirectional directional effects, splitting MBR into four components—hypothesis-to-reference likelihood, reference-to-hypothesis likelihood, and corresponding priors—enabling unified interpretation, metric/task analysis, and channel weighting for potential performance gains.","Noisy-Channel Minimum Bayes Risk Decoding  \nYusuke Sakai 1 Hidetaka Kamigaito 1 Taro Watanabe 1  \narXiv :2607 .05 198v 1 [ cs .LG] 6 Jul 2026  \nAbstract  \nMinimum Bayes Risk (MBR) decoding yields more robust and higher-quality text generation than maximum a posteriori (MAP) decoding by selecting hypotheses that maximize expected utility over sampled pseudo-references. However, there exists a discrepancy in the design: hypothesis selection calculates expected utility scores conditioned on given pseudo-references, while commonly used evaluation metrics, e.g., BLEU and COMET, are asymmetric. Therefore, it is important to consider both hypothesis-to-reference and reference-to-hypothesis directional effects. In this study, we introduce a noisy channel decomposition of MBR decoding that naturally incorporates bidirectional effects to account for these asymmetries. We decompose MBR decoding into four interacting components: hypothesis-to-reference likelihood, reference-to-hypothesis likelihood, hypothesis prior, and reference prior. This decomposition provides a unified interpretation of existing MBR variants and enables metric-and taskspecific interpretability by isolating the contribution of each channel. Our comprehensive analysis reveals that channel-wise contributions exhibit distinct characteristics across metrics while remaining consistent across tasks, and suggests that appropriate channel weighting may lead to improvements over original MBR decoding.  \n1. Introduction  \nDecoding in text generation systems typically relies on heuristic strategies such as greedy decoding, beam search, and sampling methods, which approximate maximum aposteriori (MAP) decoding by selecting the most probable hypothesis under the model’s output distribution. However, MAP decoding often fails to align with downstream evaluation metrics and human preferences (Koehn & Knowles, 2017 ; Eikema & Aziz, 2020) . Minimum Bayes Risk (MBR)  \n1Nara Institute of Science and Technology, Nara, Japan. Correspondence to: Yusuke Sakai \u003C[sakai.yusuke.sr9@is.naist.jp](sakai.yusuke.sr9@is.naist.jp) >.  \nPreprint. July 7, 2026.  \ndecoding (Goel & Byrne, 2000 ; Kumar & Byrne, 2004 ; Eikema & Aziz, 2020) yields more robust and higher-quality text generation than MAP decoding by directly optimizing expected utility with respect to an evaluation metric, such as BLEU (Papineni et al., 2002) or BERTScore (Zhang et al., 2020) . Rather than selecting the most likely hypothesis, MBR decoding selects the hypothesis that maximizes expected utility over a set of sampled pseudo-references.  \nHowever, the hypothesis selection formulation of original MBR decoding exhibits a fundamental mismatch with commonly used evaluation metrics, such as BLEU. These metrics are inherently asymmetric and do not treat hypotheses and references interchangeably. Their scores also depend on the direction of comparison. Consequently, effective hypothesis selection should consider not only reference-tohypothesis relationships, but also hypothesis-to-reference interactions, as well as priors over hypotheses and references. Such directional asymmetries and prior effects are not explicitly captured in existing MBR formulations, highlighting a fundamental limitation of the MBR decoding.  \nIn this study, we introduce a noisy-channel-based decomposition of MBR decoding that naturally accounts for such bidirectional interactions. Our formulation decomposes MBR decoding into four interacting components: the hypothesisto-reference likelihood, the reference-to-hypothesis likelihood, the hypothesis prior, and the reference prior. This decomposition provides a probabilistic interpretation that explicitly models bidirectional effects between hypotheses and references, thereby flexibly aligning the expected utility. Furthermore, this decomposition enables fine-grained interpretability. By isolating the contribution of each channel, our framework allows us to analyze which components are most influential for a given met","cbCaigwMDFTG5haX","https://ap.wps.com/l/cbCaigwMDFTG5haX","pdf",636607,1,15,"English","en",105,"# Abstract\n# Introduction\n# Background and Related Work","[{\"question\":\"How does MBR decoding differ from MAP decoding in text generation?\",\"answer\":\"MBR selects the hypothesis that maximizes expected utility over sampled pseudo-references, whereas MAP chooses the hypothesis with the highest model probability. This makes MBR directly optimize against an evaluation metric’s notion of utility.\"},{\"question\":\"Why is there a mismatch between original MBR decoding and metrics like BLEU?\",\"answer\":\"Original MBR conditions expected utility on sampled pseudo-references, but evaluation metrics such as BLEU are asymmetric and depend on comparison direction. As a result, hypothesis-to-reference and reference-to-hypothesis effects are not treated interchangeably.\"},{\"question\":\"What are the four components in the proposed noisy-channel decomposition?\",\"answer\":\"The approach decomposes MBR into hypothesis-to-reference likelihood, reference-to-hypothesis likelihood, hypothesis prior, and reference prior. This explicitly models bidirectional interactions between hypotheses and references and supports interpretability via channel-wise contributions.\"}]",1784198328,38,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":27},"noisy-channel-minimum-bayes-risk-decoding","",{"@graph":35,"@context":85},[36,53,68],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/noisy-channel-minimum-bayes-risk-decoding/84799/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-17","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},"How does MBR decoding differ from MAP decoding in text generation?","Question",{"text":75,"@type":76},"MBR selects the hypothesis that maximizes expected utility over sampled pseudo-references, whereas MAP chooses the hypothesis with the highest model probability. This makes MBR directly optimize against an evaluation metric’s notion of utility.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"Why is there a mismatch between original MBR decoding and metrics like BLEU?",{"text":80,"@type":76},"Original MBR conditions expected utility on sampled pseudo-references, but evaluation metrics such as BLEU are asymmetric and depend on comparison direction. As a result, hypothesis-to-reference and reference-to-hypothesis effects are not treated interchangeably.",{"name":82,"@type":73,"acceptedAnswer":83},"What are the four components in the proposed noisy-channel decomposition?",{"text":84,"@type":76},"The approach decomposes MBR into hypothesis-to-reference likelihood, reference-to-hypothesis likelihood, hypothesis prior, and reference prior. This explicitly models bidirectional interactions between hypotheses and references and supports interpretability via channel-wise contributions.","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":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":45,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":45,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":45,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":45,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":45,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":45,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":45,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]