[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-81662-en":3,"doc-seo-81662-105":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},81662,16904993612988,"Olivia Brown","https://ap-avatar.wpscdn.com/davatar_a8503ba1806abce46bf441b54a3ca4cd",6,"Technology","DriveMA: Driving Vision-Language-Action Models with Verifiable Meta-Actions","Driving Vision-Language-Action Models (Driving VLAs) use language to improve end-to-end autonomous driving, yet a persistent language-action gap prevents language-side decisions from being faithfully executed by predicted trajectories. DriveMA introduces a verifiable meta-action interface that summarizes future ego motion into compact language-domain intentions. Expert trajectories enable trajectory-grounded annotations, while rule-based projection verifies generated trajectories against stated intents. Action-centric supervised training and turn-level credit assignment reinforcement learning explicitly align high-level decisions with low-level planning via dense rewards.","arXiv :2605 .31271v2 [ cs .CV] 10 Jul 2026  \nDriveMA: Driving Vision-Language-Action Models with verifiable Meta-Actions  \nWeicheng Zheng 1 ,2 Yixin Huang 1 ,3 Qiao Sun 1 Derun Li 1 Hang Zhao 1 ,2 ∗  \n1 Shanghai Qi Zhi Institute 2IIIS Tsinghua University 3Tongji University  \nAbstract: Driving Vision-Language-Action Models (Driving VLAs) aim to use language to improve end-to-end planning, but the language-action gap limits this promise. We propose DriveMA, a Driving VLA framework built on verifiable meta-actions, which summarize future ego motion into compact language-domain intentions and can be constructed from expert trajectories with a trajectorygrounded annotation pipeline and can be verified against generated trajectories through rule-based projection. DriveMA exploits this verifiability with actioncentric supervised training and a data-efficient turn-level credit assignment reinforcement learning framework, explicitly aligning high-level decisions with lowlevel trajectory planning through dense rewards and precise credit assignment.  \nDriveMA sets a new state of the art on the Waymo Open Dataset Vision-based E2E Driving, achieving a Rater Feedback Score of 8.060 with a 2B model and further improving it to 8.079 with a 4B model; it also obtains competitive closedloop planning performance on NAVSIM. These results show that even a simple meta-action interface can achieve state-of-the-art planning when made verifiable and optimized for language-action alignment. Code, data, and models are available at [https://tsinghua-mars-lab.github.io/DriveMA](https://tsinghua-mars-lab.github.io/DriveMA).  \nKeywords: Autonomous Driving, Vision-Language-Action, Meta-Action  \n1 Introduction  \nDriving Vision-Language-Action Models (Driving VLAs) have recently emerged as a promising paradigm for end-to-end autonomous driving [1, 2, 3, 4] . By introducing language into the perception-to-action pipeline, these models aim to leverage semantic knowledge to improve downstream planning. Ideally, language should not merely describe the scene, but should expose highlevel driving intentions that guide low-level trajectory generation. However, the persistent languageaction gap in Driving VLAs means that plausible language-side decisions may not be faithfully executed by predicted trajectories, preventing language from becoming a truly actionable planning interface [5, 6] . Fig. 1 illustrates a concrete language-action mismatch: despite predicting acceleration at a green light, an SFT model remains nearly stationary under a near-static motion history.  \n---------------------------------------SFT--------------------------------------  \npredicted action: accelerate, straight  \n[0.00, 0.00], [0.00, 0.00], [0.00, 0.00], [0.00, 0.00], [0.43, 0.00] trajectory-implied action: stop, straight  \n--------------------------RL w Consistency Reward-----------------------  \npredicted action: accelerate, straight [0.01, 0.00], [0.71, -0.00], [2.94, -0.01], [6.44, -0.02], [10.95, -0.01] trajectory-implied action: accelerate, straight  \nFigure 1: Motivating example of the language-action gap. At a green light with near-static motion history, the SFT-only DriveMA predicts a plausible high-level decision to accelerate, but generates an almost stationary trajectory, revealing a language-action mismatch. With our RL-based explicit alignment, the trajectory better executes the predicted acceleration decision.  \n∗ Corresponding author: [hangzhao@mail.tsinghua.edu.cn](hangzhao@mail.tsinghua.edu.cn)  \nWe argue that bridging this gap requires the intermediate language interface to be verifiable. A verifiable interface allows high-level language decisions to be supervised from expert behavior, checked against generated trajectories, and optimized through explicit alignment. In this paper, we study meta-action as a simple and effective instance of such a verifiable language interface. A meta-action compactly represents the planning-relevant intent of future ego motion, such as acce","cbCaiuuZ6Y6brL3U","https://ap.wps.com/l/cbCaiuuZ6Y6brL3U","pdf",8683894,1,19,"English","en",105,"# Introduction\n## Language-action gap and verifiable interfaces\n## DriveMA framework and components\n# Experiments and results\n## Waymo Open Dataset (WOD-E2E)\n## NAVSIM and ablations","[{\"question\":\"What problem does DriveMA address in Driving Vision-Language-Action Models?\",\"answer\":\"DriveMA addresses the language-action gap, where plausible language-side decisions are not reliably executed by the generated trajectories. It motivates a need for an intermediate language interface that can be verified against motion predictions.\"},{\"question\":\"What is a meta-action in the DriveMA framework?\",\"answer\":\"A meta-action compactly represents the planning-relevant intent of future ego motion, such as accelerating, decelerating, turning, or changing lanes. It serves as a verifiable language interface between visual inputs and trajectory generation.\"},{\"question\":\"How does DriveMA verify whether a stated decision matches the generated trajectory?\",\"answer\":\"DriveMA projects trajectories back into a verifiable action space using rule-based projection, enabling checks for consistency between the high-level meta-action and the low-level predicted motion. This verifiability is then exploited during training for alignment.\"}]",1784175270,48,{"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},"drivema-driving-vision-language-action-models-with-verifiable-meta-actions","",{"@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/technology/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/drivema-driving-vision-language-action-models-with-verifiable-meta-actions/81662/",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 DriveMA address in Driving Vision-Language-Action Models?","Question",{"text":75,"@type":76},"DriveMA addresses the language-action gap, where plausible language-side decisions are not reliably executed by the generated trajectories. It motivates a need for an intermediate language interface that can be verified against motion predictions.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"What is a meta-action in the DriveMA framework?",{"text":80,"@type":76},"A meta-action compactly represents the planning-relevant intent of future ego motion, such as accelerating, decelerating, turning, or changing lanes. It serves as a verifiable language interface between visual inputs and trajectory generation.",{"name":82,"@type":73,"acceptedAnswer":83},"How does DriveMA verify whether a stated decision matches the generated trajectory?",{"text":84,"@type":76},"DriveMA projects trajectories back into a verifiable action space using rule-based projection, enabling checks for consistency between the high-level meta-action and the low-level predicted motion. This verifiability is then exploited during training for alignment.","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},[93,97,101,105,110,113,118,123,128,131,135],{"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":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":111,"slug":112},50,"technology",{"id":114,"doc_module":4,"doc_module_name":45,"category_name":115,"show_sort_weight":116,"slug":117},7,"Healthcare",40,"healthcare",{"id":119,"doc_module":4,"doc_module_name":45,"category_name":120,"show_sort_weight":121,"slug":122},8,"Research & Report",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":21,"doc_module":4,"doc_module_name":45,"category_name":136,"show_sort_weight":106,"slug":137},"General","general"]