Yandex Publishes 'The KV Cache as an Agent Runtime': Inference-Time Interactivity for LLMs

arpit_bhayani · x · 2026-09-07

Yandex Research published the full write-up behind the recent KV-cache buzz: 'The KV cache as an agent runtime.' Interactive models must take in new information mid-computation, revise trajectories, emit partial actions, and coordinate processes running at different rates — unavoidable for games, robots, live video, and OSes. While Wan-Streamer and Thinking Machines' Interaction Models solve this via retraining, new data formats, or streaming encoders, Yandex explores how much interactivity can be achieved purely at inference time by sharing and scheduling KV-cache state, letting pretrained LLMs observe, reason, and act concurrently without any training.

Related event: Yandex Turns KV Cache into an Agent Runtime for Real-Time LLM Interaction(3 posts)→

Original post →

More from Infra

Infra channel →