Chinese researchers are testing a new way to use large language models in military drones: not as chatbots, but as part of a control system for aerial combat decisions.
The system combines large language models with reinforcement learning. According to a report in the South China Morning Post, it reportedly outperforms both traditional pure reinforcement learning methods and skilled human operators when making aerial combat decisions.
How the drone AI system works
The reported design links several layers of AI processing together. First, reinforcement learning analyzes sensor data from the drone. That information is converted into vector parameters, then translated into human language so the language model can process it.
After the language model makes a decision, the system translates that decision back into specific control commands for the drone. The source describes this as a multistep process that can adapt an attack strategy up to ten times per second.
One example given is the control of an electronic warfare jammer. In that case, the system could direct the jammer to create false targets on enemy radar displays.
The important point is the role assigned to the language model. Instead of simply generating text for a human reader, the model becomes part of a loop that turns battlefield data into machine actions. Reinforcement learning handles the structured sensor analysis, while the language model is used to interpret the situation and choose among possible responses.
What the researchers used to train it
The researchers trained the language model on a wide collection of military information. The source says this included technical literature on radar and electronic warfare.
The training set also included classified materials, such as combat logs, weapons inventories, and military instruction manuals. Most of that material was in Chinese.
That training detail matters because the system is not described as a general-purpose model dropped into a drone. It was prepared with domain-specific military material, including sources connected to weapons, radar, electronic warfare, and operational records.
The South China Morning Post reports that the findings were published in the Journal of Detection & Control on October 24. The project is still in the testing phase.
Who is behind the project
The work is described as a collaboration between the Chengdu Aircraft Design Institute and Northwestern Polytechnical University in Xian.
The Chengdu Aircraft Design Institute is known for developing China's J-20 stealth fighter. Northwestern Polytechnical University in Xian is the other named partner in the project.
Because the project remains in testing, the source does not present it as an operational battlefield system. The reported significance is that researchers are experimenting with a control architecture in which an LLM can be used inside a military drone decision loop.
Part of a wider race in military AI drones
The Chinese experiment sits inside a broader international push toward military AI applications, especially in drone technology. The source points to several parallel developments involving the US military, US defense contractor Anduril, and the war in Ukraine.
The US military has invested about a billion dollars in the Switchblade kamikaze drone series. It is also developing AI pilots for both drones and fighter jets, and US forces are testing ground-based autonomous systems.
Anduril recently unveiled an AI-controlled attack drone that can autonomously carry out deadly strikes on command. The source also says the war in Ukraine has accelerated the development of autonomous, lethal unmanned systems.
According to Robert Brovdi, founder and commander of the special drone unit "Madyar Birds," Ukrainian drones could fly completely autonomously without human pilots in as little as six months. Some drones have reportedly launched AI-driven attacks on their own after losing contact with operators.
Why this development matters
The reported Chinese system shows one path military AI research may take: combining reinforcement learning, language models, sensor feeds, and control commands into one decision-making pipeline.
That pipeline is significant because it connects model reasoning to physical action. A drone's sensors produce data, the system restructures that data for the model, the model selects an action, and the action is translated into drone controls.
The source does not say the system is deployed. It does show how researchers are experimenting with LLMs beyond text generation, placing them inside systems built for electronic warfare and aerial combat decisions.
For military drones, that shift changes the role of AI from advisory software to a component that can help decide and execute actions. The reported testing phase, the military training data, and the comparison with human operators all point to a central trend: autonomous systems are becoming a major focus in drone warfare research.