LLMs are General Asynchronous Agents
The paper argues one LLM framework can handle overlapping inputs and tasks without custom training for each case.
The authors frame current agent loops as too sequential for settings where new input arrives mid-task. Their framework lets inference coroutines run with overlapping memory states, so agents can manage different kinds of concurrency. They report Qwen 3.x models operating asynchronously on streaming video understanding, videogames, and monitoring examples. Source: HF Daily Papers' note.
The authors frame current agent loops as too sequential for settings where new input arrives mid-task. Their framework lets inference coroutines run with overlapping memory states, so agents can manage different kinds of concurrency. They report Qwen 3.x models operating asynchronously on streaming video understanding, videogames, and monitoring examples. Source: HF Daily Papers' note.
score 5