When Do LLM Agents Help? Deadline-Aware Mixed-Criticality Task Scheduling at the Autonomous-Vehicle Edge

Self-driving cars do not just drive; they constantly think. To navigate a busy street, a vehicle must process a flood of data from cameras and sensors, identifying pedestrians, other cars, and road signs in a fraction of a second. Because the car's own computer cannot handle this immense workload alone, it sends these urgent calculations to nearby servers located at the edge of the network, a setup known as mobile edge computing. The challenge for engineers is that not all tasks are created equal. Some jobs, like detecting an obstacle that could cause a crash, are life-or-death and must be finished before a strict deadline. Other jobs, such as updating a map or playing music, are important but can wait if the system is busy. The goal is to build a scheduler, a digital traffic controller, that protects the life-or-death tasks without wasting the system's power, all while the car is moving and the network conditions are changing.

For years, researchers have tried to use artificial intelligence to solve this scheduling puzzle. Recently, a new type of AI called a large language model has emerged as a potential solution. These models are famous for their ability to reason and explain their choices in plain language, leading many to hope they could act as flexible, intelligent managers for complex systems. However, these models are also slow and expensive to run. The question facing the scientific community was whether adding this slow, intelligent layer to a fast, critical system would actually help, or if it would just slow everything down without providing any real benefit. A researcher set out to answer this by building a rigorous test to see exactly when, and if, these AI managers improve the performance of self-driving car networks.

The researcher constructed a detailed simulation of a network where self-driving cars offload their thinking tasks to nearby servers. They created sixty different scenarios, each with two hundred tasks arriving at random times. In these scenarios, sixty percent of the tasks were life-critical, while the rest were standard, lower-priority jobs. The researcher first built a very strong, traditional scheduling system. This system worked by grouping incoming tasks into small batches and then sorting them so that the most urgent, life-critical jobs were always considered first. It then assigned each job to the server that could finish it the soonest. This method proved to be incredibly effective, successfully completing over ninety percent of the critical tasks. It performed better than fifteen other standard methods and came within thirteen percent of the theoretical best possible performance, a benchmark calculated by a powerful offline solver that had the luxury of seeing the entire future before making any decisions.

With this high-performing baseline established, the researcher introduced the large language model. They designed a system where the AI did not make every single decision, which would have been too slow. Instead, the AI acted as a supervisor that checked in only occasionally to adjust the rules of the game. In the first part of their study, the traffic flow was steady and predictable. Under these calm conditions, the AI supervisor added no value. The traditional, rule-based scheduler was already so good that the AI could not find any extra room for improvement. The researcher found that the AI's reasoning, the auction-like bidding process, and the ability to adapt in real time were all unnecessary when the workload was stable. The simple, fast rules were sufficient to handle the traffic efficiently.

The story changed when the researcher introduced a sudden, chaotic event. They simulated a scenario where a large group of cars arrived at a busy intersection at the same time, creating a sudden surge of life-critical tasks that overwhelmed the standard scheduler. In this moment of non-stationary chaos, the traditional system began to struggle, missing deadlines it would have otherwise met. Here, the large language model control layer finally showed its worth. By reviewing the situation and adjusting the scheduling rules on the fly, the AI system managed to recover some of the lost performance, completing significantly more critical tasks than the static system or even a different type of adaptive algorithm. The AI did not solve the problem perfectly, but it found a way to squeeze out a small, statistically significant improvement that the fixed rules could not achieve.

The study concludes that the value of using these advanced AI models is entirely conditional. They are not a magic upgrade that makes everything better all the time. Instead, they earn their cost only when the environment becomes unpredictable and a fixed set of rules can no longer keep up. In steady conditions, the simple, fast methods are superior because they are efficient and reliable. But when a sudden surge of critical events occurs, the ability of the AI to reason through the chaos and adapt its strategy provides a genuine, measurable benefit. The researcher also measured the time it took for the AI to make its decisions, finding that it was far too slow to be used for every single task, confirming that it must remain a high-level supervisor rather than a direct worker. The final message for engineers is clear: do not adopt these complex AI layers by default. They are a specialized tool, useful only when the workload is volatile enough to create opportunities that a fixed policy cannot see.