Yandex Makes Pretrained LLMs Interactive by Reorganizing KV Cache, No Retraining Needed

arpit_bhayani · x · 2026-09-07

Yandex Research turned the KV cache from a mere speed optimization into an agent runtime. By letting multiple readers consume the same shared memory in different orders — no weight changes — pretrained LLMs can observe, reason, and act concurrently. A demo agent played Doom in real time, processing a constant frame stream, despite never being trained on the game. The takeaway: interactivity can come from rearranging how a model's memory is organized and exposed, a far cheaper lever than retraining.

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

Original post →

More from Infra

Infra channel →