Author | Ben Yi, Editor | De Xin
The Next-Generation Architecture Battle in China's Autonomous Driving
World models, VLA, and native multimodality... These three terms have been frequently heard at new car and autonomous driving launches over the past year. The Chinese public discourse has divided into several camps: Huawei and NIO are in the world model camp, while XPeng and Li Auto are in the VLA camp.
However, at CVPR 2026 this year, an oral paper from a Chinese automaker focuses precisely on world models. Its first paragraph, however, begins with an admission: Although the academic community has yet to form a unified and explicit definition of world models, the ability to predict the evolution of physical environments is consistently regarded as its core capability.
This paper comes from Li Auto. The pioneers at the forefront of this field have not yet drawn clear boundaries for it, yet the public discourse has long drawn the lines for them.
Simply attributing this phenomenon to "naming confusion" might be somewhat hasty. To be precise, the boundaries between these technologies were never as clearly defined as the public narrative suggests.
A Company Talking About VLA, with Two Internal Teams Working on World Models
XPeng's situation is even more interesting. It uniformly promotes VLA externally, but the five technical reports published on arXiv in the first half of 2026 are authored by four teams: Predictive World Model (PWM), Generative World Model (GWM), AI Infrastructure (AI INFRA), and Foundation Model. The acknowledgments in X-Foresight also mention that the Generative World Model team's work on X-World served as the foundation for the pre-training of its diffusion renderer.
A company that externally promotes VLA has two internal teams working on world models simultaneously.
"Native multimodality" was originally intended to end the naming confusion, but now it has branched into two distinct paths itself.
Li Auto unifies language tokens and action tokens into a single discrete codebook, making no distinction between the two from the very beginning; XPeng, on the other hand, has designed a native multimodal tokenizer, integrating different modalities early on as information enters the model.
The same Chinese term conceals two entirely different processing approaches behind it.
Calling Each Other Out in Papers, Yet the Debate Is Not About the Technical Route
If we put together the papers published by several companies in 2026, we will find that the clashes over academic and technical routes are much more intense than what is visible at launch events.
DriveLaW, a collaboration between Xiaomi Auto and Huazhong University of Science and Technology, directly points out Epona and DriveVLA-W0 in its introduction, arguing that such methods train generation and planning separately, leaving a gap between visual imagination and action selection.
The affiliated institution for DriveVLA-W0 is precisely Yinwang, the company spun off from Huawei's Car BU.
DynVLA, a collaboration between Yinwang and the Institute of Automation, Chinese Academy of Sciences (CAS), criticizes visual chain-of-thought for bringing massive redundancy and computational overhead due to pixel-level generation, thus choosing to represent the future using a "dynamics chain-of-thought."
Visual chain-of-thought, however, is precisely the core proposition of XPeng's X-Mind.
XPeng did not hold back either. X-Mind argues that placing a world model before the policy model "will bring unacceptable vehicle-side latency"; and attaching future image reconstruction at the end of the network for auxiliary supervision means "the supervision signal is hard to effectively backpropagate to the deep layers of the language model backbone, making the model prone to learning shortcuts rather than true physical reasoning."
This latter training approach is highly similar to the training paradigm proposed by Huawei's DriveVLA-W0, which uses future image prediction to provide dense supervision.
Li Auto's SparseWorld-TC, on the other hand, explicitly bypasses VAE discrete tokens and BEV (Bird's-Eye View) intermediate representations, arguing that the explicit geometric constraints of BEV would limit the flexible interaction of spatiotemporal features. The "abstract sketches" used by XPeng's X-Mind to preview the future, however, are fundamentally based on BEV.
Looking at it this way, what these four companies criticize is exactly "how the other side does it specifically." None of them truly denies the act of "predicting the future," meaning no one says the other side has taken the wrong path. X-Mind even explicitly states that integrating predictive world models into autonomous driving has become an industry consensus.
Therefore, the question "VLA or world model?" might be asking the wrong thing entirely.
The next round of competition in autonomous driving after end-to-end is not a binary opposition of either A or B.
The Real Divergence: How Much Detail Is Needed to Preview the Future?
When humans drive, they also "think" it through in their minds first.
Seeing a car in the front right crossing the lane line, we quickly judge: if it merges over, should I slow down or yield? Humans just don't render these imagined scenes into a 4K video in their minds. That thought usually only contains the car's outline, the relative positions of each other, and the possible actions.
Making machines drive like humans requires doing similar things. The biggest difference between humans and machines, however, lies in: how much detail should be retained in this generated "thought."
XPeng is betting on "high abstraction."
The technical report for X-Foresight argues that "world representation should maintain high abstraction; excessive visual details in the latent space will dilute the model's structured world understanding capability." X-Mind further compresses the future 12 frames into 96 tokens to express BEV layout, navigation intent, and traffic rules; realistic rendering is handed over to another module and does not directly participate in reasoning.
Xiaomi has chosen the opposite direction. In the paper on "DriveLaW," the authors argue that existing approaches "do not treat the generator's internal latent variables as planning states, leaving a gap between visual imagination and action selection." Therefore, DriveLaW advocates that "the internal representations learned by the video generator from massive driving videos are inherently valuable, and the planner should directly use these representations to avoid disconnecting generation and planning."
Li Auto walks a middle path between the two. Zhan Kun, head of Li Auto's foundation model, explained it clearly in his keynote speech at GTC 2026: "If we directly generate future real images, the computational cost would be extremely high, so we choose to make predictions in the latent space." Li Auto calls this capability "multimodal thinking," meaning the model imagines the future in the latent space rather than drawing out real images.
Huawei's DynVLA compresses it even more aggressively, using a dedicated tokenizer to compress future evolution into a small set of discrete tokens, and splitting ego-vehicle dynamics and environmental dynamics into two separate parts for independent modeling.
On the matter of "predicting the future," the four companies have provided four different answers.
This divergence in technical paths may allow people to see more clearly what kind of theoretical foundation leading companies have actually applied in the next-generation intelligent driving technology architecture, compared to the binary theory of "VLA or world model."
There is also another comparison that has yet to reach a conclusion: XPeng uses BEV as the baseboard for the abstract world, while Li Auto considers BEV to be a limitation. Whether it is an effective geometric prior or a structural constraint that hinders model learning may require more mass-production results to answer.
The Route Has Not Yet Been Decided, But They Are First Tripped Up by the Same Problem
Many users have encountered a specific type of problem after turning on the intelligent driving system: the car can drive, but it takes the wrong route.
In April 2026, Fudan University, together with Tsinghua AIR, the Institute of Automation of the Chinese Academy of Sciences, and DiDi, conducted a set of experiments: deleting the correct navigation information received by the end-to-end model, or even replacing it with random steering commands, resulted in almost no drop in comprehensive planning scores, and sometimes even a slight improvement.
At least in the tested models, although navigation was integrated, it did not truly affect planning. The models rely more on the immediate road structure, camera feeds, and common driving patterns in the training data.
The reason is not complicated. What end-to-end models receive are often just discrete commands like "turn left," "go straight," or "turn right." A human driver knows they need to turn right in a few hundred meters and will select the right lane in advance; the command the model receives at this moment, however, might still be "go straight." During training, it sees a human changing lanes but may not know whether it is to overtake or for an upcoming right turn.
The causal relationship between navigation and driving actions is broken.
This problem is frequently seen in real-world tests by many media outlets.
When a media outlet tested Horizon Robotics' HSD in Tianjin, they encountered a complex intersection that required going straight through the intersection first and then entering the central left-turn lane. HSD 1.6 had a probability of mistakenly entering the U-turn lane and deviating from the path; V2.0 got it right in both reproductions under heavy rain conditions.
Another media outlet, while testing the IM LS8 equipped with Momenta's solution, had only 3 manual interventions over hundreds of kilometers, yet two of them occurred during the transition between main and auxiliary roads: the system failed to locate the intersection accurately and took the wrong route.
Our colleagues at Dasuanli also encountered a situation during a test drive of DeepRoute's VLA version where the vehicle first changed lanes to the left, and after realizing the route was wrong, continuously crossed two solid lines to change back. Previously, during a test drive of Li Auto's VLA, the team also recorded similar issues.
In the papers, each company debates highly detailed technical issues such as "how to predict the world." However, in real-world tests of mass-produced intelligent driving products, these systems have been successively tripped up by the same navigation challenge.
Yet, in the technical architecture stories told externally by the vast majority of companies, this problem has hardly been discussed head-on.
Beyond the Models, There Are Two Even Harder Problems to Solve
The first problem is whether Large Language Models (LLMs) can run efficiently in vehicles.
Ren Shaoqing, head of NIO's intelligent driving, provided the judgment that: under the same computational load, the memory bandwidth demand of Transformers is 8 to 70 times that of CNNs; the key difference lies not only in computing power but more so in memory bandwidth. A single NIO Shengji NX9031 chip has a bandwidth exceeding 500 GB/s.
Li Auto, in its Mach M100 chip paper, emphasizes that hardware-software co-design needs to solve not just computation, but more critically, how data is transported across time and space. The M100 has a peak bandwidth of 273 GB/s and significantly reduces reliance on multi-level caches through a dataflow architecture.
These two figures cannot be simply used to judge the strength of the chips, but they illustrate the same problem: as large models enter vehicles, the competition has shifted from "how much can be computed" to "whether data can be delivered to where it needs to go in a timely manner."
Horizon Robotics has submitted a different kind of answer. Among its 11 accepted papers at CVPR 2026, 4 focus on 3D reconstruction, 3 on robot embodiment, 2 on perception, along with 1 on end-to-end planning and 1 participating in world model evaluation; there are 0 papers directly related to building world models or VLA.
The only end-to-end planning paper, ResAD, also does not have the model predict trajectories from scratch; instead, it first uses a constant velocity model to calculate an inertial reference trajectory, and then has the model learn the deviation relative to it. While others are packing more capabilities into large models, Horizon Robotics returns the parts that can be solved by physical formulas back to physical formulas.
This is consistent with what Su Qing, Vice President and Chief Architect at Horizon Robotics, said in a public speech late last year: "My personal judgment is that there is a high probability that the next three years will be about extreme optimization on existing systems, rather than a reconstruction of the theoretical core. So everyone shouldn't be too hyped; we are entering a phase of hard times again."
Whether he is sober or conservative cannot be concluded yet. But at least, Horizon Robotics has made a technical choice that keeps a distance from industry buzzwords.
The second problem is whether capabilities can be verified by external parties.
World models play a key role in each company's narrative: as closed-loop simulation environments and engines for online reinforcement learning. Zhan Kun was very direct at GTC, stating, "The capability of the World Simulator actually directly determines the upper limit of reinforcement learning."
So how high is this upper limit right now?
Another oral paper at CVPR this year, WorldLens, is a benchmark specifically for evaluating driving world models. The numbers it measured are: in closed-loop simulation, letting a pre-trained planner drive in the generated "world," the route completion rate is at most 13.51% and at least 6.89%. The route completion rate here refers to how much of the preset route is completed before termination due to collision, driving off the road, or timeout. The open-loop performance of the same batch of models is 71% to 79%, which drops to 4.8% to 10.6% in closed-loop.
The exact words in the paper are "all methods collapse under closed-loop conditions."
But a disclaimer is needed here: WorldLens tested exclusively open-source academic models, excluding any production-level models from any company. This result cannot be extrapolated to the mass-production models of any automaker.
However, there is still a question worth asking: world models are described by various companies as a crucial foundation for closed-loop simulation and reinforcement learning, yet this search did not find any mass-production model that has submitted a public, horizontally comparable closed-loop evaluation. What the outside world sees remains primarily each company's own datasets and metrics.
According to public papers and research findings found by Dasuanli, Li Auto, Xiaomi, Huawei, and Horizon Robotics have used public benchmarks such as nuScenes and NAVSIM; the core results disclosed by XPeng, NIO, Momenta, Zhuoyu, and DeepRoute mainly come from self-built datasets or lack horizontally comparable public metrics. Xiaomi even specifically noted that its results did not use reinforcement learning post-training and reward model post-processing.
Therefore, from the perspective of combining theory with practice, the results of these benchmark tests cannot indicate who has stronger mass-production capabilities. When each company uses different test papers and scoring methods, it is actually very difficult for the outside world to judge just how large the leading margin claimed at launch events really is.
This gap also appears in Momenta.
In Chinese context communication, Momenta is often referred to as the "first Physical AI stock." However, in its IPO documents submitted to the Hong Kong Stock Exchange, Physical AI does not appear, and world model appears only once, corresponding to the R7 which is "under development." The market-facing identity label, when translated into legal documents for investors, ultimately becomes just a project under research and development.
The Solution to the Next Problem, Might Not Be in the Algorithms at All
When a media outlet tested XPeng's VLA 2.0 in Germany, they encountered a bicycle towing a small trailer. However, the test vehicle could only follow behind the entire time, with no intention to overtake. The German Road Traffic Regulations require that when motor vehicles overtake cyclists in urban areas, the lateral safety distance must not be less than 1.5 meters.
The long-held perception in China is: the messier the road conditions and the more complex the dynamic game, the stronger the model performance, so if it can drive in Guangzhou, it can drive in Munich. The dynamic game in Munich is not as complex as in Guangzhou, but compliance is a different kind of difficulty. The model not only needs to "know how to drive," but also needs to accurately understand local regulations and implement the rules into every single action.
According to existing plans, the EU urban advanced driver assistance system regulation UN R171 Series 02 is expected to take effect by the end of 2026. XPeng hopes to be among the first to get approved.
Over the past year, companies in the industry have spent a lot of energy debating which route they are taking.
The answer found in each company's papers is: the general direction is converging, and the differences are hidden in future representation, vehicle-side efficiency, and training methods. The results of real-world tests, however, show that more specific issues such as navigation, public evaluation, and local compliance are far from being resolved.
VLA, world models, native multimodality... the names can continue to change. What ultimately decides the winner is whether the car can take the right route.
References
The technical judgments in this article are primarily based on the following public papers and materials. All papers can be searched for their original text on arXiv.
Li Auto
SparseWorld-TC: Trajectory-Conditioned Sparse Occupancy World Model
CVPR 2026 (Oral) | arXiv:2511.22039
https://arxiv.org/abs/2511.22039
M100: An Orchestrated Dataflow Architecture Powering General AI Computing
arXiv:2604.17862
https://arxiv.org/abs/2604.17862
Zhan Kun, "MindVLA-o1: Unlocking the Omni-Paradigm," NVIDIA GTC 2026 Keynote, 2026-03-17
XPeng
X-Foresight: A Joint Vision-Action Causal Forecasting Network via Predictive World Modeling
arXiv:2605.24892
https://arxiv.org/abs/2605.24892
X-Mind: Efficient Visual Chain-of-Thought via Predictive World Model for End-to-End Driving
arXiv:2606.28758
https://arxiv.org/abs/2606.28758
Xiaomi Auto
DriveLaW: Unifying Planning and Video Generation in a Latent Driving World
CVPR 2026 | arXiv:2512.23421
https://arxiv.org/abs/2512.23421
Huawei (Authored by Yinwang Intelligent, Noah's Ark Lab, etc.)
DriveVLA-W0: World Models Amplify Data Scaling Law in Autonomous Driving
ICLR 2026 | arXiv:2510.12796
https://arxiv.org/abs/2510.12796
DynVLA: Learning World Dynamics for Action Reasoning in Autonomous Driving
ICML 2026 | arXiv:2603.11041
https://arxiv.org/abs/2603.11041
Horizon Robotics
ResAD: Normalized Residual Trajectory Modeling for End-to-End Autonomous Driving
arXiv:2510.08562
https://arxiv.org/abs/2510.08562
Summary of Horizon Robotics' Accepted Papers at CVPR 2026 (Source for the distribution of 11 research directions)
https://www.horizon.auto/news/technology/439
Evaluation Benchmarks and Third-Party Research
WorldLens: Full-Spectrum Evaluations of Driving World Models in Real World
CVPR 2026 (Oral) | arXiv:2512.10958
https://arxiv.org/abs/2512.10958
Unveiling the Surprising Efficacy of Navigation Understanding in End-to-End Autonomous Driving
Fudan University, Tsinghua University AIR, Institute of Automation of the Chinese Academy of Sciences, DiDi | arXiv:2604.12208
https://arxiv.org/abs/2604.12208
Epona: Autoregressive Diffusion World Model for Autonomous Driving
arXiv:2506.24113
https://arxiv.org/abs/2506.24113
(Appears in the text as the target of explicit criticism in Xiaomi's paper)
Other Public Materials
Momenta Global Limited Global Offering Document, The Stock Exchange of Hong Kong Limited, Stock Code 6880
Public interview with NIO's Ren Shaoqing (Statements related to memory bandwidth)
German Road Traffic Regulations (StVO) Section 5, Paragraph 4; UN R171 Series 02