[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-86250-en":3,"doc-seo-86250-105":29,"detail-sidebar-cat-0-en-105":90},{"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":11,"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},86250,687197207919,"Theodora","https://ap-avatar.wpscdn.com/avatar/a000253d6f5f7c60be?x-image-process=image/resize,m_fixed,w_180,h_180&k=1779446848396160552",8,"Research & Report","Comparative Analysis of GAT and BERT for Human-Like Playtesting","Accurate modeling of player experience is essential for designing engaging puzzle games, motivating data-driven automated playtesting that predicts human-like decisions. Existing approaches often require extensive feature engineering and continual model adjustments when new mechanics appear. To reduce maintenance needs, the work compares generalized transformer-based BERT and graph attention network GAT architectures for relationally representing Candy Crush Saga board states. Experiments assess move prediction and difficulty alignment against CNN baselines, highlighting stronger performance on challenging configurations.","Comparative Analysis of GAT and BERT for Human-Like Playtesting  \nKleio Fragkedaki, Theodoros Panagiotakopoulos, Matteo Biasielli, Hui Wang  \nAI Center of Excellence  \nKing  \nStockholm, Sweden  \n{claire.fragkedaki, theodoros.panagiotakopoulos, matteo.biasielli, [maddy.hui.wang](maddy.hui.wang}@king.com)[}](maddy.hui.wang}@king.com)[@king.com](maddy.hui.wang}@king.com)  \narXiv :2607 . 1 150 1v 1 [ cs .AI] 13 Jul 2026  \nAbstract—Accurately modeling and understanding player experience is crucial for designing engaging puzzle games. To achieve this, a common approach involves collecting diverse user data to train predictive playtesting models that mimic player behavior. However, existing data-driven methods often lack the ability to capture the full range of player strategies and require extensive feature engineering and network architecture modeling. This limitation becomes particularly evident when new game mechanics or features are introduced, which necessitate continual adjustments to the models. To address these challenges, we propose a more generalized representation that reduces —or even eliminates — the need for ongoing feature-engineering maintenance. Specifically, we investigate two general-purpose network architectures: (a) a transformer-based model (BERT) and (b) a graph attention model (GAT), both of which are designed to effectively capture the relational structure of Candy Crush Saga (CCS) game boards. Our experiments compare these approaches to Convolutional Neural Networks (CNN) baselines, revealing better performance on challenging board configurationsand underscoring the benefits of our generalizable representation.  \nI. INTRODUCTION  \nIn game development, balancing difficulty is crucial for player retention and engagement. Challenge plays a fundamental role in player experience, and research indicates that players quickly abandon games that fail to provide an appropriate level of difficulty [5] . To optimize gameplay, developers traditionally rely on playtesting to evaluate new levels and finetune game parameters before release [18] . However, human playtesting is costly, time-consuming, and often fails to capture the diversity of player behaviors [2] . As a result, machine learning-powered automated playtesting has become a scalable and efficient solution, offering a rapid, data-driven analysis of gameplay across various player profiles and play styles [7, 8] . In this framework, human-like behavior refers to the ability of a model to replicate player decisions. This is evaluated using two criteria: (a) move prediction accuracy — how often the model’s actions match human choices in historical data [19, 16] — and (b) level difficulty alignment — how closely model performance correlates with aggregated player success rates [7] .  \nIn grid-based puzzle games like Candy Crush Saga (CCS), players generally swap items on a two-dimensional board to create matches based on specific patterns, aiming to accomplish predefined objectives. Due to the spatial layout of  \nthese games, Convolutional Neural Networks (CNNs) [13] are a popular choice for modeling player interactions as they excel at extracting features from grid-like data [4, 19], and have been shown to be computationally efficient in game playtesting compared to methods like Monte-Carlo Tree Search (MCTS) [7] . However, CNN architectures often struggle to generalize when the board size changes or new game mechanics are introduced, partly due to the convolutional architecture [3] and partly because their grid-based representation does not reflect the underlying relational topology of the level [11] .  \nTo address these challenges, we explore two alternative deep learning architectures for automated playtesting of Candy Crush Saga levels: (a) a Bidirectional Encoder Representation from Transformers (BERT) [6] and (b) a Graph Attention Network (GAT) [23] . The BERT model, commonly used in natural language processing, is adapted to interpret game states as text-bas","cbCaicwUzEFrAzBM","https://ap.wps.com/l/cbCaicwUzEFrAzBM","pdf",4018318,3,1,"English","en",105,"# Introduction\n## Human-like playtesting and evaluation criteria\n## Why CNNs face generalization limits\n## Motivation for BERT and GAT\n# About Candy Crush Saga","[{\"question\":\"What does “human-like playtesting” mean in this study?\",\"answer\":\"It refers to a model replicating player decisions using move prediction accuracy and level difficulty alignment based on aggregated player success rates.\"},{\"question\":\"Why do CNN-based playtesting models struggle when the game changes?\",\"answer\":\"CNNs generalize poorly when board size varies or new mechanics are introduced, partly due to the convolutional architecture and partly because grid representations do not capture underlying relational topology.\"},{\"question\":\"How are BERT and GAT adapted for Candy Crush Saga board modeling?\",\"answer\":\"BERT is adapted by treating game states as text-based inputs to provide flexible board configuration handling, while GAT uses graph representations to capture relational dependencies and connections between game elements.\"}]",1784209811,20,{"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":85,"head_meta":87,"extra_data":89,"updated_unix":27},"comparative-analysis-of-gat-and-bert-for-human-like-playtesting","",{"@graph":35,"@context":84},[36,52,67],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,49],{"item":40,"name":41,"@type":42,"position":21},"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":20},"https://docshare.wps.com/document/research-report/",{"item":50,"name":13,"@type":42,"position":51},"https://docshare.wps.com/document/comparative-analysis-of-gat-and-bert-for-human-like-playtesting/86250/",4,{"url":50,"name":13,"@type":53,"author":54,"headline":13,"publisher":56,"fileFormat":59,"inLanguage":23,"description":14,"dateModified":60,"datePublished":61,"encodingFormat":59,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":55},"Person",{"url":40,"name":57,"@type":58},"DocShare","Organization","application/pdf","2026-07-26","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":20},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"What does “human-like playtesting” mean in this study?","Question",{"text":74,"@type":75},"It refers to a model replicating player decisions using move prediction accuracy and level difficulty alignment based on aggregated player success rates.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"Why do CNN-based playtesting models struggle when the game changes?",{"text":79,"@type":75},"CNNs generalize poorly when board size varies or new mechanics are introduced, partly due to the convolutional architecture and partly because grid representations do not capture underlying relational topology.",{"name":81,"@type":72,"acceptedAnswer":82},"How are BERT and GAT adapted for Candy Crush Saga board modeling?",{"text":83,"@type":75},"BERT is adapted by treating game states as text-based inputs to provide flexible board configuration handling, while GAT uses graph representations to capture relational dependencies and connections between game elements.","https://schema.org",{"og:url":50,"og:type":86,"og:title":13,"og:site_name":57,"og:description":14},"article",{"robots":88,"canonical":50},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":91},[92,96,100,104,109,114,119,122,126,129,133],{"id":21,"doc_module":4,"doc_module_name":45,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":51,"doc_module":4,"doc_module_name":45,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":105,"doc_module":4,"doc_module_name":45,"category_name":106,"show_sort_weight":107,"slug":108},5,"Comic",60,"comic",{"id":110,"doc_module":4,"doc_module_name":45,"category_name":111,"show_sort_weight":112,"slug":113},6,"Technology",50,"technology",{"id":115,"doc_module":4,"doc_module_name":45,"category_name":116,"show_sort_weight":117,"slug":118},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":120,"slug":121},30,"research-report",{"id":123,"doc_module":4,"doc_module_name":45,"category_name":124,"show_sort_weight":28,"slug":125},9,"Religion & Spirituality","religion-spirituality",{"id":28,"doc_module":4,"doc_module_name":45,"category_name":127,"show_sort_weight":28,"slug":128},"World Cup","world-cup",{"id":130,"doc_module":4,"doc_module_name":45,"category_name":131,"show_sort_weight":130,"slug":132},10,"Lifestyle","lifestyle",{"id":134,"doc_module":4,"doc_module_name":45,"category_name":135,"show_sort_weight":105,"slug":136},19,"General","general"]