[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-83317-en":3,"doc-seo-83317-105":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},83317,1374391974585,"Genevieve","https://ap-avatar.wpscdn.com/davatar_276721f389ce27ea32af1340a28f341c",8,"Research & Report","Post-Training in End-to-End Autonomous Driving: A Unified View","End-to-end models directly map multimodal inputs to future trajectories, powering recent autonomous driving research through joint perception–prediction–planning. However, safety-critical, interaction-rich environments expose weaknesses of classic offline open-loop imitation: small execution errors compound over time, recovery behaviors are rare in data, and long-horizon objectives like safety and driving comfort cannot be captured by pointwise labels. This survey unifies post-training for autonomous driving by defining its scope and organizing literature into four supervision families, detailing capabilities, limitations, and open challenges to enable reliable, efficient refinement beyond imitation.","arXiv :2607 .08072v 1 [ cs .CV] 9 Jul 2026  \nPost-Training in End-to-End Autonomous Driving: A Unified View  \nRuining Yang 1 ⋆ , Muxing Wang 1 ⋆ , Yixiao Chen 1 , Tongfei Guo 1 , Yi Xu 1 , Can Cui2 , Zichong Yang2 , Yitian Zhang 1 , Ziran Wang2 , Yun Fu 1 , and Lili Su 1  \n1 Northeastern University, Boston, USA  \n2 Purdue University, West Lafayette, USA  \nAbstract. End-to-end models that map multimodal inputs directly to future trajectories/maneuvers have emerged as an increasingly prominent research paradigm in autonomous driving. This class of models includes both Vision-Language-Action models and trajectory-generative planners. Unlike classic machine learning applications, autonomous vehicles operate in safety-critical and interaction-intensive environments where traditional open-loop imitation of expert demonstrations is not sufficient to ensure reliability. In particular, small execution errors can accumulate over time, while recovery behaviors are scarce in training data. In addition, long-horizon objectives such as safety and driving comfort are not captured by pointwise labels either. These limitations have motivated a shift toward post-training techniques, which further refine driving policies beyond pure imitation. This survey presents a unified view of post-training for autonomous driving by defining its scope and organizing the existing literature into four major families based on the form of supervision they use. For each family, we discuss its capabilities, limitations, and open challenges. We aim to facilitate a systematic understanding of this emerging area and stimulate future research on reliable and efficient post-training for autonomous driving.  \nKeywords: Autonomous Driving · Post-training · Reinforcement Learning · Distillation · Preference Alignment  \n1 Introduction  \nEnd-to-end models have become an increasingly prominent research paradigm in autonomous driving, wherein a single network takes multimodal inputs (e.g. camera images, navigation commands, and the vehicle’s own state) and directly outputs future trajectories that instruct the maneuvers of the ego vehicle [19, 25 , 44] . Their growing popularity is driven by the promise of jointly optimizing perception, prediction, and planning, reducing reliance on handengineered intermediate interfaces, learning richer cross-modal representations, and exhibiting stronger adaptability to complex driving scenarios.  \n⋆ Equal contribution. Email: [yang.ruini@northeastern.edu](yang.ruini@northeastern.edu) ,  \n[wang.muxin@northeastern.edu](wang.muxin@northeastern.edu)  \n2 R. Yang, M. Wang et al.  \n⋮ ⋮  \nFig. 1: Two-stage training paradigm for autonomous driving policies. Stage 1 builds initial driving competence through imitation learning, while Stage 2 further improves the policy through post-training.  \nRepresentative end-to-end driving models have evolved from early behavioralcloning approaches that map raw pixels from front-facing camera and high-level navigation commands (e.g. turn left/right) directly to low-level controls [5, 10], to trajectory-generative planners that predict future trajectories [9, 22 , 30 , 72 , 78 , 96], and Vision-Language-Action (VLA) models that leverage advanced multimodal representations and language reasoning to improve scene understanding and decision making [46, 85 , 86 , 87] . Despite the architectural differences among these approaches, they largely rely on imitation learning (i.e., supervised learning on offline training data) to produce coarse actions or trajectories that mimic expert driving behavior from demonstrations.  \nUnlike classic machine learning applications [80, 83], autonomous vehicles operate in safety-critical and interaction-intensive environments where traditional open-loop imitation is not sufficient to ensure reliable driving behaviors. In closed-loop executions, where a model’s actions influence future observations, even small errors can compound over time; for example, as shown in Fig. 1 , a minor lane","cbCaiiHXcFYUY8I1","https://ap.wps.com/l/cbCaiiHXcFYUY8I1","pdf",1105974,1,24,"English","en",105,"# Introduction\n# Post-Training Definition and Scope\n## Distillation Family\n## Preference-Based Alignment Family\n## Reinforcement Learning Family\n## Test-Time Refinement Family\n# Evaluation Protocols and Challenges\n# Open Problems and Future Directions","[{\"question\":\"Why is post-training needed for end-to-end autonomous driving instead of relying only on offline imitation learning?\",\"answer\":\"Offline open-loop imitation cannot guarantee reliability in safety-critical, interaction-intensive environments. Small execution errors can accumulate in closed-loop rollouts, and important driving qualities are not well reflected by pointwise imitation losses.\"},{\"question\":\"How does post-training differ from initial training and learning from scratch?\",\"answer\":\"Post-training is treated as a distinct refinement stage applied beyond initial imitation competence. It leverages additional supervision signals to further improve driving policies rather than learning the full behavior from scratch.\"},{\"question\":\"What are the four major families of post-training methods organized in the survey?\",\"answer\":\"The survey organizes post-training literature into four families based on the form of supervision used: distillation, preference-based alignment, reinforcement learning, and test-time refinement.\"}]",1784186698,60,{"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},"post-training-in-end-to-end-autonomous-driving-a-unified-view","",{"@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/post-training-in-end-to-end-autonomous-driving-a-unified-view/83317/",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},"Why is post-training needed for end-to-end autonomous driving instead of relying only on offline imitation learning?","Question",{"text":75,"@type":76},"Offline open-loop imitation cannot guarantee reliability in safety-critical, interaction-intensive environments. Small execution errors can accumulate in closed-loop rollouts, and important driving qualities are not well reflected by pointwise imitation losses.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does post-training differ from initial training and learning from scratch?",{"text":80,"@type":76},"Post-training is treated as a distinct refinement stage applied beyond initial imitation competence. It leverages additional supervision signals to further improve driving policies rather than learning the full behavior from scratch.",{"name":82,"@type":73,"acceptedAnswer":83},"What are the four major families of post-training methods organized in the survey?",{"text":84,"@type":76},"The survey organizes post-training literature into four families based on the form of supervision used: distillation, preference-based alignment, reinforcement learning, and test-time refinement.","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},[93,97,101,105,109,114,119,122,127,130,134],{"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":28,"slug":108},5,"Comic","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":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":45,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":131,"doc_module":4,"doc_module_name":45,"category_name":132,"show_sort_weight":131,"slug":133},10,"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":45,"category_name":136,"show_sort_weight":106,"slug":137},19,"General","general"]