From late 2025 to early 2026, the hottest concepts in the intelligent driving industry were undoubtedly world models and VLA (Vision-Language-Action). A new model architecture or a new technical route was enough for companies to showcase their technological prowess.
But recently, these two concepts seem to be gradually cooling down.
Rather than discussing what new models have been launched, more and more companies are beginning to discuss how models enter training, simulation, and verification processes, and how they ultimately translate into actual capabilities in mass-produced vehicles. This is not because technological innovation has slowed down, but rather that competition in intelligent driving is entering a more difficult phase.
As world models and VLA become mandatory options for realizing intelligent driving, the next challenge for the industry is: Do these models actually make cars drive better?
01. Why New Models No Longer Equal New Capabilities?
In recent years, new technical routes have continuously emerged in the intelligent driving industry. End-to-end approaches changed the development method of traditional modular systems, world models allowed systems to begin understanding and inferring dynamic environments, and VLA further explored the connection between visual understanding and driving decisions.
When a technology first emerges, the model itself is part of the competitiveness. Whoever proposes a new architecture first, whoever demonstrates new capabilities first, gains more attention.
However, as more and more companies enter the same technical route, it becomes difficult to directly equate model architecture with actual driving capability.
A company using a world model does not mean its vehicle is necessarily better at handling complex intersections; a VLA system being able to understand more semantic information does not mean this information has stably translated into reasonable driving actions.
The model is only the starting point of R&D. From model training to data supplementation, simulation testing, real-vehicle validation, and mass production deployment, every step in between affects the final result.
In April this year, when XPeng released the X-World world model technical report, world models had already been used for the training, data generation, and closed-loop verification of the second-generation VLA.
Compared to solely showcasing what the model can generate, this shift aligns more closely with current industry needs. For R&D teams, whether a model can predict the future is just the first step.
More importantly, can it be used to identify problems, construct more test scenarios, and help judge exactly what changes occurred in the vehicle's driving performance after a model update?
As more companies possess similar technical directions, comparing model names and architectures alone makes it difficult to determine whose actual capabilities are stronger.
Ultimately, everything comes back to the car.
02. How to Prove That the Model Has Improved?
After large models entered the intelligent driving industry, the real question faced is how to prove that the model is sufficiently intelligent.
Traditional intelligent driving systems consist of multiple modules such as perception, prediction, planning, and control. When an issue arises in a specific link, R&D personnel can relatively clearly analyze where the problem lies and adjust the corresponding module.
The development of end-to-end models changed this R&D approach.
More and more capabilities are integrated into a unified model and training system. Changes brought by a single training session may simultaneously affect the vehicle's behavior in multiple scenarios.
To improve navigation capabilities at complex intersections, R&D teams can only supplement more relevant data and retrain the model. After the update, the vehicle's performance at these intersections indeed improved, but the overall behavior of the model may also change. Will originally stable scenarios be affected? Will the vehicle's risk judgment become more conservative or aggressive? Will optimizing for a local problem introduce new issues in other scenarios?
This is an increasingly obvious difference between large-model-based intelligent driving and traditional software development.
Adding a new feature in traditional software allows for relatively independent testing. However, capabilities within driving models are interconnected, making it difficult to simply split an update into several non-interfering parts. Therefore, iterating on intelligent driving models cannot just look at whether new capabilities have been added; one must also judge whether the entire system has become better after the update.
A vehicle successfully passing through a complex construction zone once only proves that the model has the capability to handle that specific scenario.
Once truly entering mass production, what matters more is whether the vehicle can maintain reasonable judgment when the position of construction barriers changes, surrounding traffic increases, or even the road environment is completely different.
Solving one problem without creating new ones becomes harder as the model grows more complex.
In the future, the gap between companies may not just lie in whether the model learns new capabilities, but also in whether the company has sufficiently effective tools and systems to know exactly what changed after each model update.
This is why simulation and verification are becoming increasingly important in the R&D of next-generation intelligent driving.
03. World Models Allow Problems to Be Verified Repeatedly
Real-world road testing is the most effective way to discover problems, but it is not suitable for solving all R&D issues. If a test vehicle shows hesitation at a construction zone today, even if the problem is resolved later, the position of barriers, surrounding vehicles, and road environment may be completely different the next time it passes through.
For some low-frequency scenarios, companies may need to wait a long time to encounter similar problems again. This means that after discovering a problem on real roads, R&D personnel cannot immediately obtain enough samples to determine whether the model made a random error or has a more universal capability defect. The emergence of world models provides a new R&D approach.
After a vehicle discovers a problem on real roads, R&D teams no longer have to wait for similar scenarios to reappear; instead, they can proactively add variations around this problem.
What happens if the construction area expands? Does the original decision still hold if a vehicle suddenly cuts in from the surroundings? Will the model hesitate again after road boundaries change?
In this way, a problem accidentally discovered on real roads can gradually expand into a set of scenarios that can be tested repeatedly.
Real roads remain the ultimate environment for verifying intelligent driving capabilities, and world models can hardly fully cover all complex and random changes in the real world. However, world models can change the way R&D personnel handle problems.
Now, companies can start from a real problem, proactively change surrounding conditions, and observe the model's performance under different circumstances. For increasingly complex driving models, this capability will become more important. Because what intelligent driving truly needs to solve is never a fixed scenario, but a class of constantly changing problems.
04. The Next Stage Competes on the Ability for Continuous Progress
The declining hype around world models and VLA does not mean these technologies are less important. On the contrary, as they begin to enter training, simulation, and verification workflows, competition enters the true stage of implementation.
In the future, different companies may adopt similar model architectures and possess increasingly comparable computing power and training data. But final driving performance may still show significant gaps.
These gaps stem from R&D systems outside the model itself.
If a company discovers a problem with the vehicle on real roads and can only wait for more data accumulation before training and road testing, the speed of model improvement will be limited by real-road constraints.
If another company can quickly locate problems, construct more varied scenarios around them, verify different model performances through simulation, and then confirm on real roads, the entire iteration process could be faster.
From this perspective, the next stage of intelligent driving competition is shifting from individual models to the entire R&D closed loop.
Who can discover problems faster, who can expand problems more effectively, and who can accurately judge changes after model updates will all affect the speed of vehicle capability improvement.
In the past, the industry found it easier to compare who released a new model first or whose model scale was larger.
Going forward, competition will focus on: What specific problem did a model update solve? Can new capabilities be stably repeated in more scenarios? Will solving one problem bring new issues in other scenarios? Can these changes eventually stay stably in mass-produced vehicles?
These questions are less topical than a new model name, but they determine whether technology can ultimately become product capability.
World models and VLA will continue to evolve, and new model architectures will keep emerging. But as more technologies enter the same R&D system, the leading advantage gained by releasing a single new technology may become increasingly insufficient.
This is perhaps why world models and VLA have not been frequent promotional highlights in recent months compared to previous periods.