Meta researchers show Qwen 3.x can run as general asynchronous agents without task-specific training

yresearch · hf · 2026-09-30

A new paper, "LLMs are General Asynchronous Agents," argues that modern LLM agents are stuck in sequential read-think-act loops while real use cases like voice assistants, embodied agents, and monitoring need to ingest new input mid-thought. The authors build a general asynchronous LLM framework where users (or agents themselves) define inference coroutines with overlapping memory states, unifying specialized architectures (voice, video streaming, async tool calls) under one paradigm. They show Qwen 3.x models handle streaming video understanding, videogames, and monitoring asynchronously — with no task-specific training.

Original post →

More from AGI Musings

AGI Musings channel →