When driving, you have probably encountered vehicles with a small blue indicator light on. These vehicles generally have driving assistance or autonomous driving features activated.
Many drivers tend to keep their distance from such vehicles, while some others may accelerate to cut in. In many cases, autonomous vehicles choose to decelerate and yield.
Why do autonomous vehicles still yield even when they clearly detect a cut-in?
This phenomenon occurs because the way autonomous driving systems judge cut-in vehicles differs from that of human drivers.
01. What Does a Cut-in Look Like to an Autonomous Driving System?
For an autonomous driving system, a cut-in is not understood as a simple action, but rather as a continuously changing state of a traffic participant.
When a vehicle first appears in an adjacent lane, the system needs to determine its position, speed, heading, and lateral movement trend.
If the other vehicle gradually approaches the ego lane, the system must continue to assess whether it is merely slightly crossing the lane line or preparing for an actual lane change. How much distance remains between it and the ego vehicle? Given their current speeds, will a conflict occur in the next few seconds?
This highlights the critical role of the prediction module in autonomous driving.
The prediction module does not simply guess where the vehicle will be in the next second. Instead, based on the vehicle's position and speed changes over a past period, combined with lane markings, road structure, and the states of surrounding vehicles, it generates the potential future trajectories of the other vehicle.
During the upgrade to its second-generation VLA, Xpeng publicly mentioned that for sudden scenarios such as the leading vehicle braking to a halt, pedestrians turning back, and adjacent vehicles forcefully cutting in, the system introduces faster streaming auto-regressive reasoning and predicts the future possible behaviors of surrounding traffic participants through a world model.
Its X-Foresight predictive world model can deduce the movement trends of surrounding traffic participants within the next 6 seconds, assisting the system in anticipating scenario changes such as cut-ins, pedestrian direction changes, and vehicle braking in advance. Meanwhile, the Infini-VLA long-sequence architecture can retain road environment data from the past 30 seconds, linking fragmented road events to provide a longer-term scenario context for autonomous driving decisions.
This is particularly critical for cut-in judgment. A cut-in is generally a dynamic process lasting several seconds. Long-sequence memory helps the system more accurately judge the evolution of the other vehicle's intentions, rather than relying solely on instantaneous frames for decision-making.
The technical route of Huawei's Qiankun ADS 5 also reflects a similar approach. Its officially released vehicle-side world behavior model introduces the safety risk field theory for the first time, making real-time decisions through risk heat maps, which can reduce collision risks by 50%. Huawei describes this as moving further from merely identifying targets to judging risks.
In addition, Huawei's cloud-based world model introduces a multi-agent game mechanism for the first time, increasing training intensity by 10 times, which is directly related to game-theoretic decision-making in cut-in scenarios.
Qiankun OS also reduces in-vehicle signal latency by 30% through technologies such as a deterministic scheduling engine, ensuring the timely execution of low-latency tasks like emergency cut-ins.
Therefore, when a vehicle just begins to approach the ego lane, the autonomous driving system does not wait until it has fully encroached before making a decision. It may have already anticipated the potential conflict of this lane change in advance based on the other vehicle's lateral speed, distance, and movement trend.
A simple cut-in maneuver, which a human might perceive as not yet having fully cut in, may already be seen by the autonomous driving system as a trajectory approaching the ego vehicle, requiring proactive deceleration to respond.
02. Why Do Autonomous Vehicles Still Choose to Yield After Detecting a Cut-in?
For an autonomous driving system, once it determines that the other vehicle is cutting in, the real difficulty lies in what to do next.
Assuming an adjacent vehicle is cutting into the ego lane, the autonomous driving system has at least several handling options: maintaining speed to let the other vehicle complete the lane change; decelerating appropriately to proactively leave space for it; or, if conditions permit, adjusting its own trajectory to avoid a conflict.
While it may seem like a simple matter of whether to yield or not, it actually involves a continuous decision-making process.
The system needs to simultaneously consider the ego vehicle's speed, the distance to the vehicles ahead and behind, the speed and lateral movement of the target vehicle, vehicles in adjacent lanes, and the current braking capability of the ego vehicle.
This is especially true on congested urban roads, where the distance between vehicles is already relatively small, and a single maneuver can easily affect the vehicles behind.
Therefore, autonomous driving systems generally do not treat holding their position as the sole objective; instead, they prioritize avoiding uncontrollable conflicts.
This is why some autonomous driving systems exhibit behaviors that human drivers might perceive as overly polite.
For a human driver, holding ground slightly might just be a normal driving interaction; for an autonomous driving system, if the other party's behavior has high uncertainty, proactively reducing speed makes it easier to keep risks within a controllable range.
In its previously disclosed autonomous driving technologies, Li Auto had already specifically optimized congested game-theoretic lane changes. The official release notes for its OTA 5.2.0 explicitly mentioned optimizing lane change timing, lane change space, and strategies for game-theoretic waiting during lane changes, reducing the waiting time for initiating lane changes and overtaking, and improving the decisiveness of lane change waiting decisions and lane change effectiveness. At the same time, it adjusted the speed control strategy for scenarios where other vehicles cut in, optimizing the speed control strategy for car-following and cut-in scenarios, reducing jerkiness and hard braking during car-following, and enhancing car-following comfort and a sense of security.
This indicates that Li Auto's technical direction focuses more on making more decisive decisions in game-theoretic scenarios rather than simply yielding. In its publicly released technical materials in 2024, yielding to approaching adjacent vehicles was included in the user manual, and the OTA 6.0.0 release notes mentioned navigating around obstacles and improving the response to dynamic targets. This demonstrates that a cut-in is not simple obstacle avoidance, but an interactive decision-making process between vehicles.
From the perspective of game theory, a cut-in is essentially a non-cooperative game, where the cutter gains time benefits by violating rules, while the cut-in victim is forced to bear the risks of deceleration and safety hazards.
A paper published in 2026 in Transportation Research Part C: Emerging Technologies proposed a hierarchical FSM-Nash framework. It decouples the decision-making process into a strategic layer (game-theoretic coordination, achieving multi-stage behavior transitions through a finite state machine) and an execution layer (a constrained Nash equilibrium planner), achieving a 50% improvement in safety and an 11% improvement in efficiency in dense traffic cut-in scenarios.
The study also introduced a self-evolving Bayesian risk mechanism, dynamically calibrating risk thresholds through real-time interaction data and regional driving patterns, enabling the system to adapt to driving style differences across various regions.
If the autonomous driving system only detects a vehicle ahead, it is prone to making sudden braking maneuvers; if it can further determine that an adjacent vehicle is cutting in at a relatively fast speed, the system can adjust its speed earlier, turning a sudden cut-in into a relatively smooth speed change.
03. What Autonomous Driving Really Needs to Solve Is the Degree of Yielding
For the cut-in scenario, what the autonomous driving system needs to do is not to switch from always yielding to never yielding.
A truly mature strategy should know when to yield, how much to yield, and when there is no need to yield.
In the Xmart OS 4.3.0 version pushed by Xpeng in July 2023, XNGP officially incorporated an anti-cut-in strategy, officially described as dynamically controlling the following distance in congested scenarios while optimizing car-following capabilities. The same version also added different driving style options for the lane change strategy and optimized lane change timing and execution methods.
The underlying logic of such technologies is that if the following distance is always kept very large, adjacent vehicles can more easily find space to cut in; if the following distance is compressed too small, it may lead to frequent braking of the ego vehicle, or even increase the risk of collision.
Therefore, true anti-cut-in is not simply about reducing the following distance, but enabling the vehicle to dynamically adjust its behavior based on traffic conditions.
In normal traffic flow, an adjacent vehicle merely approaching does not necessarily mean it will definitely cut in, so there is no need to decelerate suddenly at this time. When the other vehicle has clearly cut into the ego lane, the system can decelerate in advance to reduce conflicts. If the other vehicle has completed the cut-in, the system should quickly restore a reasonable following distance rather than continuously maintaining an excessively large safety space.
This series of actions, in a traditional modular architecture, involves at least target tracking, trajectory prediction, intention judgment, risk assessment, longitudinal control, and lateral planning. The difficulty lies in the fact that these modules cannot make decisions independently.
If prediction suggests the other vehicle might cut in, planning must prepare in advance; if planning decides to yield, control must execute it with smooth deceleration; if the other vehicle ultimately does not cut in, the system must promptly correct its judgment.
In an end-to-end architecture, these processes are no longer passed serially as independent modules but are unified into a single neural network for joint inference. Perception features directly participate in the generation of planning decisions, and the system's response to cut-ins is completed synchronously within a continuous inference flow, no longer relying on explicit intermediate variables such as prediction results.
In other words, end-to-end does not eliminate the judgment process in the traditional modular architecture; rather, it couples them in parallel within an implicit feature space, thereby reducing information loss between modules and internalizing the objective trade-offs that require cross-module negotiation in the traditional architecture into a unified optimization process.
This is also why more and more automakers are now starting to specifically optimize cut-ins, congested lane changes, and approaching adjacent vehicles.
A review of human-machine mixed traffic research published in the China Journal of Highway and Transport in 2026 pointed out that human-driven vehicles exhibit adaptive behaviors when facing autonomous vehicles. This change affects car-following and lane-changing behaviors and is closely related to subjective factors such as the driver's trust in autonomous driving technology and driving style, as well as objective factors such as the penetration rate of autonomous vehicles and the driving environment.
Another independent study on AV (autonomous vehicle) penetration rate and driving style further found that the impact of penetration rate varies from person to person.
Aggressive and normal drivers exhibit more aggressive decision-making behaviors when the AV penetration rate increases, while the behaviors of defensive drivers are unaffected by changes in the penetration rate.
This means that the impact of an increased AV penetration rate on human-driven vehicle behavior is not a linear "the higher, the more significant," but is closely related to the individual driving style of the driver.
The conservative behavior of autonomous driving itself may actually have a negative impact on traffic flow.
A study using co-simulation analysis found that the impact of conservative driving varies with speed limits, mainly because the hard deceleration of autonomous vehicles leads to a sharp reduction in the safety distance to the following vehicle.
Another traffic flow simulation study pointed out that autonomous vehicles are more conservative than human drivers in terms of acceleration behavior and lateral movement, requiring larger lane change gaps and forcing them to decelerate and wait more frequently.
This provides a possible explanation for why autonomous vehicles are more easily cut off in congested scenarios, as larger safety gaps can become one of the triggers for being cut off.
In high-density mixed traffic, it is extremely rare to simultaneously satisfy the three dimensions of safety, interaction quality, and traffic performance. Although excessive yielding by autonomous driving improves safety in certain metrics, it may incur hidden costs in other dimensions, such as traffic efficiency and the experience of following vehicles.
Therefore, the perception of human drivers that autonomous vehicles are always yielding, to some extent, precisely illustrates that when autonomous vehicles enter a real human-machine mixed environment, the system needs to predict not only the road but also other drivers.
What autonomous driving truly needs to improve in the future is not simply becoming tougher, but handling cut-in behaviors more precisely—yielding in advance when it should, and not over-yielding when it shouldn't. When faced with forced cut-ins, it should also be able to maintain a reasonable driving rhythm while ensuring safety.
In this way, autonomous driving will not appear unintelligent due to excessive conservatism, nor will it increase new risks just to appear as assertive as human drivers.