September 23, 2026
AI agents will be able to receive new data while answering: AsyncReasoning speeds up response times by up to 12 times
On September 23, Yandex Research introduced AsyncReasoning: in extreme scenarios, the method reduced perceived latency by up to 12 times. Time to the first visible token fell by up to 80 times. The method works without additional model training.
Previously, a sequential LLM would first finish reasoning, while new camera frames or tool results waited for the next run. AsyncReasoning now separates input, Thinker's hidden reasoning, and Writer's visible response. Thinker can stop Writer, accept fresh data, and adjust the response.
No retraining. The method uses positional embeddings and KV-cache manipulation rather than fine-tuning. Yandex is developing a shared async I/O framework in SGLang and its own GPU kernels. A demo of Qwen3.5 playing Doom from a continuous video stream is also in development.
The team continues to develop a general async I/O framework in SGLang and its own GPU kernels.
Source
