[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-85666-en":3,"doc-seo-85666-105":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},85666,549758252649,"Ivy","https://ap-avatar.wpscdn.com/avatar/8000253669c5317157?_k=1778319167496531819",8,"Research & Report","Scene Graph Thinking Reinforcing Structured Visual Reasoning for Multimodal Large Language Models","Multimodal Large Language Models (MLLMs) show strong perception and reasoning, yet many systems treat visual content as isolated objects and ignore structured relationships needed for efficient target navigation in complex scenes. Scene Graph Thinking (SaGe) introduces explicit scene-graph representations for fine-grained visual reasoning. An automated data engine converts image–text corpora into hierarchical nodes and relation edges, then produces 120K high-quality reasoning traces for training. Two-stage graph-aligned post-training applies supervised fine-tuning and node-as-proxy reinforcement rewards, improving performance across eight multimodal benchmarks on detailed perception and reasoning tasks.","Scene Graph Thinking: Reinforcing Structured Visual Reasoning for Multimodal Large Language Models  \nZhiwei Yang * 1 Yuanchen Wu * 1 Nan Zhang 1 Yucong Meng 2 Ke Yan† 1 Shouhong Ding 1  \narXiv :2607 .05716v3 [ cs .CV] 13 Jul 2026  \nAbstract  \nMultimodal Large Language Models (MLLMs) have demonstrated strong perception and reasoning capabilities. However, most existing models focus on isolated objects and neglect structured relationships for efficient target navigation, limiting their performance on visually intensive tasks. To address this challenge, we introduce Scene Graph Thinking (SaGe), a novel paradigm that enables fine-grained and structured visual reasoning through explicit scene-graph representations. Specifically, we first introduce an automated data engine that converts flat image–text corpora into structured scene graphs, where hierarchical entities constitute the nodes and diverse visual relations define the edges. Building upon this, we construct 120K high-quality training data by sampling reasoning traces from scene graphs. Then two-stage graph-aligned post-training paradigms are introduced, where supervised fine-tuning internalizes MLLMs with structured reasoning, and subsequent reinforcement fine-tuning proposes node-as-proxy graph rewards to consolidate efficient graph exploration.  \nWith curated data and graph-aligned training, our approach achieves significant improvements across eight multimodal benchmarks, demonstrating strong effectiveness on fine-grained perception and reasoning tasks. Code is available at [https://github.com/zwyang6/SaGe](https://github.com/zwyang6/SaGe).  \n1. Introduction  \nMultimodal Large Language Models (MLLMs) (Hurst et al., 2024 ; Jaech et al., 2024 ; Comanici et al., 2025) have advanced rapidly in recent years, showing growing compe-  \n*Equal contribution . †Corresponding author. 1Tencent Youtu Lab, Shanghai, China 2Fudan University, Shanghai, China. Correspondence to: Ke Yan \u003C[kerwinyan@tencent.com](kerwinyan@tencent.com) >.  \nProceedings of the 43 rd International Conference on Machine Learning, Seoul, South Korea. PMLR 306, 2026 . Copyright 2026 by the author(s) .  \n(a) Existing Methods  \nQuery: What's the color of the license plate?  \n\u003Cthink> I can see a stone monument… I can see the traffic lights… I see a car… I see a yellow and metal plate.  \n\u003Canswer> The color of the license plate is yellow.  \n\u003Cthink> I can observe a white and metal car.  \nOn the front bumper of this car , i can find a yellow license plate made of sturdy metal.  \n\u003Canswer> The color of the license plate is yellow.  \nFigure 1. Our motivation. (a) Previous methods overlook structured relationships within the scene, leading to inefficient target navigation and suboptimal performance. (b) The proposed SaGe internalizes structured scene graphs to enable fine-grained visual reasoning, achieving more efficient and reliable performance.  \ntence in multimodal perception (Lai et al., 2024 ; Liu et al., 2025) understanding (Wang et al., 2023 ; Zhu et al., 2025b), and reasoning (Wang et al., 2025a ; Shen et al., 2025) . As instruction-following abilities and cross-modal transfer continue to improve, MLLMs are increasingly capable of handling a broad spectrum of tasks, driving their adoption in both general-purpose and specialized applications (Yin et al., 2024 ; Wang et al., 2025d) . However, despite these advances, current MLLMs still struggle in complex visual scenariosand exhibit suboptimal performance in fine-grained analysis. (Zhang et al., 2025a ; Wang et al., 2025b) .  \nRecent milestones, such as OpenAI’s o3 (OpenAI, 2025), pioneer to address above limitations by integrating visual information during the reasoning (Shao et al., 2024) . Following these works, several studies (Zheng et al., 2025 ; Fan et al., 2025 ; Wang et al., 2025b) have adopted a “think-withimage” paradigm. This literature typically applies cropping and zoom-in operations to magnify target subjects, thereby enhancing fine-grained perception.","cbCaihYbyYev5ZjC","https://ap.wps.com/l/cbCaihYbyYev5ZjC","pdf",19718118,1,22,"English","en",105,"# Introduction\n## Existing Methods\n## Proposed Method: Scene Graph Thinking (SaGe)\n## Data Engine and Scene Graph Construction\n## Graph-Aligned Training Approach","[{\"question\":\"What problem does Scene Graph Thinking (SaGe) address in existing MLLMs?\",\"answer\":\"Existing MLLMs often focus on isolated objects and neglect structured relationships, which harms fine-grained analysis and efficient target navigation in visually complex scenarios.\"},{\"question\":\"How does SaGe represent visual information for reasoning?\",\"answer\":\"SaGe uses explicit scene-graph representations where hierarchical entities form nodes and visual relations define edges, enabling context-aware relational reasoning and structured traversal.\"},{\"question\":\"What training strategy does SaGe use after building scene graphs?\",\"answer\":\"SaGe applies two-stage graph-aligned post-training: supervised fine-tuning to internalize structured reasoning, followed by reinforcement fine-tuning that uses node-as-proxy graph rewards to consolidate efficient graph exploration.\"}]",1784205476,55,{"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},"scene-graph-thinking-reinforcing-structured-visual-reasoning-for-multimodal-large-language-models","",{"@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/scene-graph-thinking-reinforcing-structured-visual-reasoning-for-multimodal-large-language-models/85666/",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 Scene Graph Thinking (SaGe) address in existing MLLMs?","Question",{"text":75,"@type":76},"Existing MLLMs often focus on isolated objects and neglect structured relationships, which harms fine-grained analysis and efficient target navigation in visually complex scenarios.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does SaGe represent visual information for reasoning?",{"text":80,"@type":76},"SaGe uses explicit scene-graph representations where hierarchical entities form nodes and visual relations define edges, enabling context-aware relational reasoning and structured traversal.",{"name":82,"@type":73,"acceptedAnswer":83},"What training strategy does SaGe use after building scene graphs?",{"text":84,"@type":76},"SaGe applies two-stage graph-aligned post-training: supervised fine-tuning to internalize structured reasoning, followed by reinforcement fine-tuning that uses node-as-proxy graph rewards to consolidate efficient graph exploration.","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"]