LLMs are too slow and random to control physical systems — so they only give high-level advice.

Physical systems like robots, smart grids, and multi-agent teams need exact safety and balance. Today, people want to use large language models, or LLMs, to make big decisions for them. But LLMs are slow and unpredictable, which can break physical safety rules. The authors studied how to safely add LLMs to these control setups. They suggest a two-part design. The LLM acts as a slow planner that picks broad goals. Meanwhile, a fast, proven inner controller runs directly on the hardware to keep it physically stable. Under this view, computer issues match classic network problems. For example, LLM lag acts like network delay, and AI errors act like physical disturbances. They found that as models get smarter, safety guarantees often drop.

Why it matters

As LLMs are used in self-driving cars, drones, and smart grids, their unpredictability can cause crashes or failures. This survey shows a way to keep human-like decision-making while maintaining safety, by separating high-level commands from low-level control. Without this, using LLMs in physical systems risks accidents.

Questions answered