[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-82927-en":3,"doc-seo-82927-105":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},82927,8796095461610,"Oliver","https://ap-avatar.wpscdn.com/davatar_276721f389ce27ea32af1340a28f341c",8,"Research & Report","MoP-JEPA Hard-Assigned Predictor Mixtures for Stochastic JEPA World Models","MoP-JEPA addresses stochastic JEPA world-model planning where a single regressor cannot represent multiple valid next latent states. Under stochastic transitions, squared and cosine regression each collapse diverse successors into conditional means or direction-only compromises. MoP-JEPA introduces K hard-assigned predictor heads with a context-only router to generate a finite candidate successor set in one pass. On OGBench transitions, graph search with MoP-JEPA achieves 0.85 success versus 0.02–0.09 for single-output baselines, using verified-route success to separate true successors from indiscriminate coverage.","MoP-JEPA: Hard-Assigned Predictor Mixtures  \nfor Stochastic JEPA World Models  \nZhi Song 1,2 , Ximing Xing2 , Zhenchao Tang2 , Hanbo Huang2 , Weilong Yan2 , Tianxu Lv2 , Minghao  \nYang2 , Zhongzheng Niu2 , Bing He*2 , Lusheng Wang* 1 , Jianhua Yao*2  \n1 City University of Hong Kong, China 2Tencent, China  \narXiv :2607 .05238v2 [ cs .AI] 11 Jul 2026  \nAbstract  \nJEPA world models commonly predict the next latent state with one regressor. Under stochastic transitions, squared and cosine regression return the conditional mean and its normalized direction, respectively: a single compromise that may match no valid successor. MoP-JEPA instead uses K hardassigned heads and a context-only router to produce a finite candidate set in one pass. On held-out OGBench transitions, graph search with single-output predictors succeeds on 0.02– 0.09 of queries, whereas MoP-JEPA reaches 0.85. To distinguish useful successors from indiscriminate coverage, we also measure verified-route success (realroute), which checks after graph construction whether the proposal contains a path of real transitions. MoP-JEPA leads this same-protocol metric on all three mazes; an MDN attains high raw coverage but predicts many nonexistent edges.  \nIntroduction  \nJEPA world models plan by rolling a latent predictor forward (LeCun 2022; Zhou et al. 2024; Assran et al. 2025) . They work well when the future is nearly deterministic. Stochastic dynamics violate this assumption. One context can have several valid successors, and planning requires those successors rather than their average.  \nA deterministic JEPA predictor trained by squared regression returns the conditional mean; with normalized targets and cosine loss, it returns the mean direction. A gated weighted-sum MoE still emits one vector per context and therefore retains the same single-output restriction. MoPJEPA replaces that output with a predictor containing K hard-assigned heads. Each target updates its nearest head, while a router learns which heads are active from context alone. The resulting candidate set can be used by graph search or MPC. Figure 1 summarizes the argument.  \nWe first characterize the single-output restriction of deterministic and fused-MoE predictors and relate hard assignment to conditional vector quantization (Props. 1–3) . Trained models are then examined in latent and decoded space, with additional tests on antmaze, image observations, ETH/UCY, SVHN, and DINO-WM. Finally, we determine whether the candidate successors form usable transitions for planning. Because a large candidate set can inflate coverage without learning context dependence, the evalu  \nation includes  a context-free codebook, shuffled contexts,* Corresponding author.  \nfailure fix  \n\n| Branch |  | Collapse |  | Enumerate |  | Audit |\n| --- | --- | --- | --- | --- | --- | --- |\n| one context |  | single head |  | hard heads |  | router-gated |\n| many futures | predicts mean |  | cover modes |  | realroute plans |  |\n\npredictor-level mechanism planning-level evidence  \nFigure 1: A single JEPA predictor represents a stochastic transition by one compromise output. Hard-assigned heads retain multiple candidate successors, whose transition validity is evaluated through planning rather than raw coverage alone.  \nrouter gating, transition precision, and verified-route success (Sec. ) .  \nThe multiple-choice objective itself is classical (GuzmanRivera, Batra, and Kohli 2012; Lee et al. 2016) . Our contributions are its use as a finite successor interface for JEPA world models, an empirical account of how single-output prediction affects planning under stochastic transitions, and an evaluation protocol that distinguishes context-dependent transitions from coverage obtained by indiscriminate prediction.  \nRelated Work  \nJEPA world models. I-JEPA (Assran et al. 2023) and VJEPA (Bardes et al. 2024) established latent-regression prediction. DINO-WM (Zhou et al. 2024) and V-JEPA 2 (Assran et al. 2025) extend it to action-condit","cbCaiplvV81mMM02","https://ap.wps.com/l/cbCaiplvV81mMM02","pdf",5963986,4,1,"English","en",105,"# Abstract\n# Introduction\n# Related Work","[{\"question\":\"Why do single-output JEPA predictors fail under stochastic dynamics?\",\"answer\":\"Stochastic transitions allow one context to have several valid successors, but deterministic and fused MoE predictors still output only one vector. This forces a compromise that may not match any valid successor, harming planning quality.\"},{\"question\":\"How does MoP-JEPA generate multiple candidate successors in one pass?\",\"answer\":\"MoP-JEPA replaces the single-output interface with K hard-assigned predictor heads and a context-only router. Each target updates its nearest head, while the router learns which heads activate from context to form a finite candidate set.\"},{\"question\":\"What is “verified-route success” and why is it important?\",\"answer\":\"Verified-route success checks after graph construction whether the proposed candidate set contains a path of real transitions. This distinguishes useful successors from coverage that comes from indiscriminate prediction, such as MDN achieving high raw coverage but predicting many nonexistent edges.\"}]",1784184009,20,{"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},"mop-jepa-hard-assigned-predictor-mixtures-for-stochastic-jepa-world-models","",{"@graph":35,"@context":84},[36,52,67],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"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":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":20},"https://docshare.wps.com/document/mop-jepa-hard-assigned-predictor-mixtures-for-stochastic-jepa-world-models/82927/",{"url":51,"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","2026-07-22","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},"Why do single-output JEPA predictors fail under stochastic dynamics?","Question",{"text":74,"@type":75},"Stochastic transitions allow one context to have several valid successors, but deterministic and fused MoE predictors still output only one vector. This forces a compromise that may not match any valid successor, harming planning quality.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"How does MoP-JEPA generate multiple candidate successors in one pass?",{"text":79,"@type":75},"MoP-JEPA replaces the single-output interface with K hard-assigned predictor heads and a context-only router. Each target updates its nearest head, while the router learns which heads activate from context to form a finite candidate set.",{"name":81,"@type":72,"acceptedAnswer":82},"What is “verified-route success” and why is it important?",{"text":83,"@type":75},"Verified-route success checks after graph construction whether the proposed candidate set contains a path of real transitions. This distinguishes useful successors from coverage that comes from indiscriminate prediction, such as MDN achieving high raw coverage but predicting many nonexistent edges.","https://schema.org",{"og:url":51,"og:type":86,"og:title":13,"og:site_name":57,"og:description":14},"article",{"robots":88,"canonical":51},"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":20,"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"]