Robin Kahlow

Robin Kahlow is a principal research scientist at Runway working on generative video and interactive world models. His work spans training and inference for video models, real-time interfaces and simulated environments for robotics.[1][5][2][3]

Kahlow led research on Gen-Turbo, which reduced high-quality video generation times from minutes to seconds, according to his CVMP speaker biography. He also helped build Runway’s Gen-series image and video models, from pretraining through inference optimization.[5]

For a model to create an interactive world, it must generate frames as a user watches and do so quickly enough to respond. Kahlow describes adapting a foundational audio-video model to WorldPrompt, post-training it for autoregressive generation and using distillation to reach real-time speed. WorldPrompt can steer characters and events in a scene, though he cautions that long-term memory and difficult actions remain open problems.[1]

In the Latent Space interview, Kahlow points to a possible use beyond games: giving agents many simulated environments to explore. In the system described, the agent observes generated video and audio rather than reading a structured world state.[1]

Kahlow is a co-author of Solaris, a 2026 paper on generating interactive interfaces frame by frame in response to mouse actions. Its authors combine an autoregressive visual model with a language model that interprets user intent; the paper presents the approach as research toward dynamically generated interfaces.[3]

He also co-authored Runway Robotics’ study of simulated robot policy evaluation. The team tested eight manipulation policies in a world model and compared the results with physical-world outcomes. The early experiment explores whether simulation can make parts of evaluation faster before committing to hardware tests.[2]

Kahlow’s public GitHub profile also lists projects in visual programming, geometric algebra and simulation, including NetPrints, tfga and jaxga.[4]