[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-83913-en":3,"doc-seo-83913-105":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},83913,8796095461610,"Oliver","https://ap-avatar.wpscdn.com/davatar_276721f389ce27ea32af1340a28f341c",8,"Research & Report","Claim-Level Rubric Rewards for Video Caption Reinforcement Learning","Claim-Level Rubric Rewards (CuRe) is a structured reward framework for reinforcement learning in dense video captioning, targeting the reward-design bottleneck. Prior reward designs either use holistic response-level scoring, which weakens factual accuracy and invites stylistic reward hacking, or rely on reference-based alignment that penalizes valid open-ended descriptions lacking strict textual overlap. CuRe decomposes captions into category-aware atomic claims using a rubric and verifies them at claim level. A reference-anchored calibration mechanism uses reference claim salience and visual grounding to provide bounded credit, improving both accuracy and completeness. Integrated with GRPO, CuRe improves performance across multiple video captioning and caption-to-QA tasks, with strong results for a 30B-A3B model.","arXiv :2607 .05 150v 1 [ cs .CV] 6 Jul 2026  \nClaim-Level Rubric Rewards for Video Caption Reinforcement Learning  \nMingqi Gao1,3,∗, Hongyuan Dong3,∗,†, Yifei Chen3,∗, Zhisheng Zhong3 , Zheng Ruan3 , Wenjin Hou3 , Yu Chen2 , Han Hu3,‡, Yansong Tang1,B  \n1 Tsinghua Shenzhen International Graduate School, Tsinghua University  \n2University of Chinese Academy of Sciences  \n3LLM Department, Tencent  \n[minkkigao@gmail.com](minkkigao@gmail.com) [tang.yansong@sz.tsinghua.edu.cn](tang.yansong@sz.tsinghua.edu.cn)  \nAbstract  \nIn this paper, we introduce Claim-Level Rubric Rewards (CuRe), a structured reward framework designed to address the reward-design bottleneck in reinforcement learning for dense video captioning. Existing reward designs generally fall into two categories: holistic response-level judgment across heterogeneous criteria, or alignment-based evaluation against reference captions.  \nHowever, both paradigms suffer from fundamental limitations. Holistic rewards struggle to ensure factual accuracy and are prone to stylistic reward hacking, while reference-based rewards overly rely on rigid textual alignment, failing to preserve the completeness and diversity inherent to open-ended generation tasks. To address these challenges, CuRe reformulates reward modeling as fine-grained claim-level verification. Specifically, CuRe decomposes captions into category-aware atomic claims through a structured rubric, converting holistic evaluation into simpler and more reliable claim-level verification. To reward captions that are both faithful and informative, we further introduce a reference-anchored calibration mechanism that leverages reference claims assalience anchors and visual grounding as factual evidence. This encourages reward-level mode seeking toward vital content, while assigning bounded credit to visual grounded details absent from the reference, preserving descriptiveness without drifting into verbosity. Integrated into the standard GRPO paradigm, CuRe consistently improves both accuracy and completeness. Across captioning, re-captioning, and caption-to-QA tasks, our 30B-A3B model consistently outperforms its same-scale models and surpasses substantially larger baselines on multiple benchmarks.  \n1 Introduction  \nVideo captioning represents a fundamental capability in video understanding. It serves as an essential data source for video-language alignment, a cornerstone for advanced reasoning (Maaz et al., 2024 ; Yang et al., 2024), and a critical enabler for downstream tasks ranging from controllable video generation (Xiong et al., 2024) to content recommendation. In pursuit of models capable of generating detailed and faithful descriptions, the prevailing paradigm relies on Supervised Fine-Tuning (SFT) using high-quality corpora distilled from proprietary models or human annotation (Yuan et al., 2025 ; Chen et al., 2024) . Despite its ubiquity, this approach suffers from distinct bottlenecks. Beyond the exorbitant cost of data curation, SFT inherently suffers from exposure bias (Arora et al., 2022 ; Song & Zheng, 2026): the model is trained exclusively on teacher-generated prefixes and scarcely learns to recover from its own deviations at inference, with errors compounding over dense captions. Moreover, the paradigm tends to capture the teacher’s style and confidence rather than its factual grounding (Gudibande et al., 2023) .  \nReinforcement Learning (RL) offers a viable alternative by optimizing the captioner against reward signals derived from its own rollouts. Current approaches design rewards along two axes: holistic scoring, which prompts a reward model to directly evaluate the generated text against coarse criteria (Ye et al., 2025 ; Xing et al., 2026), and reference-grounded matching, which aligns the rollout against a ground-truth reference via lexical metrics (e.g., ROUGE, CIDEr) (Pasunuru & Bansal, 2017 ; Oliveira dos Santos et al., 2021) or model judgments (Menget al., 2025 ; Zhong et al., 2025) . While these m","cbCaikGDhDH9t122","https://ap.wps.com/l/cbCaikGDhDH9t122","pdf",6083825,1,22,"English","en",105,"# Introduction\n## Supervised Fine-Tuning Bottlenecks\n## Reinforcement Learning Reward Design Limits\n## Proposed Claim-Level Rubric Rewards (CuRe)","[{\"question\":\"What problem does CuRe address in reinforcement learning for video captioning?\",\"answer\":\"CuRe targets the reward-design bottleneck by replacing coarse reward signals with fine-grained claim-level verification derived from a structured rubric.\"},{\"question\":\"How does CuRe differ from holistic response-level reward scoring?\",\"answer\":\"Holistic rewards collapse multiple visual dimensions into a single scalar, which can cause stylistic reward hacking; CuRe converts evaluation into more reliable, fine-grained claim checks for factual and informative content.\"},{\"question\":\"How does CuRe reduce issues caused by rigid reference alignment?\",\"answer\":\"CuRe introduces reference-anchored calibration that leverages reference claim salience anchors and visual grounding, giving bounded credit to visual details not present in the reference while preserving descriptiveness.\"}]",1784191411,55,{"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},"claim-level-rubric-rewards-for-video-caption-reinforcement-learning","",{"@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/claim-level-rubric-rewards-for-video-caption-reinforcement-learning/83913/",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},"What problem does CuRe address in reinforcement learning for video captioning?","Question",{"text":75,"@type":76},"CuRe targets the reward-design bottleneck by replacing coarse reward signals with fine-grained claim-level verification derived from a structured rubric.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does CuRe differ from holistic response-level reward scoring?",{"text":80,"@type":76},"Holistic rewards collapse multiple visual dimensions into a single scalar, which can cause stylistic reward hacking; CuRe converts evaluation into more reliable, fine-grained claim checks for factual and informative content.",{"name":82,"@type":73,"acceptedAnswer":83},"How does CuRe reduce issues caused by rigid reference alignment?",{"text":84,"@type":76},"CuRe introduces reference-anchored calibration that leverages reference claim salience anchors and visual grounding, giving bounded credit to visual details not present in the reference while preserving descriptiveness.","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"]