Research Scientist, Real-Time Interactivity / Inference

Real-time interactivity can come from inference-time methods applied to an existing model, from architectures designed around latency from the start, or from the interplay between the two. All three matter to us. Our customers bring diverse architectures onto our platform, and we aim to both get more out of what they've already trained and shape how the next generation of models is designed. We're hiring a Research Scientist to lead that work. You'll invent the methods that let people, agents, and robots drive video and world models frame-by-frame. Your work will shape what Reactor's customers ship, and where our platform goes next in real-time interactive world models. What You'll Do Lead a research agenda on real-time interactive generation — inference-time methods, architectural design, and where you think the field should go next Partner closely with our inference engineers to understand where the real bottlenecks in serving real-time interactive world models live, and let those challenges shape your research Work with model partner teams to build a first-hand view of the broader interactive world model landscape — the architectures, failure modes, and open problems the field is running into Publish and present your work at top venues (NeurIPS, ICLR, ICML, CVPR, ICCV, SIGGRAPH, and equivalents) Who We're Looking For PhD in ML, computer vision, graphics, robotics, or a related field, or equivalent practical experience A track record of research in one or more of: real-time or streaming video generation, autoregressive / causal video diffusion, diffusion distillation, efficient attention or state-space models for generation, or interactive controllable generation You think of latency, memory, and model quality as one problem rather than three Fluent in PyTorch and/or JAX and comfortable with large-scale training and inference infrastructure Real systems intuition — can read a profiler, reason about memory bandwidth, and have a productive conversation with the kernel engineer next to you Value quality over quantity in publishing; treat widely adopted open-source work as a mark of real impact Strong Candidates May Also Have Experience With Real-time interactive generation systems (StreamDiffusion-style, CausVid / Self Forcing-style, interactive world model demos, and similar) Action conditioning, camera control, or other structured forms of user input Writing or modifying CUDA / Triton / custom attention kernels Representative Projects Identifying a recurring failure mode across the world models running on our stack and formulating a robust, generalizable solution Proposing a new architecture, training regime, or distillation approach for real-time interactive generation that others build on Building a benchmark for long-horizon coherence under real-time interactive constraints Benefits A seat on Reactor's research team, with room to influence direction and lead your own research agenda A unique vantage point across the field — real workloads from multiple state-of-the-art model families running on one stack, with direct access to the teams building them Close collaboration with a world-class ML inference engineering team — your research ships on the stack every Reactor customer runs on, reaching people, agents, and robots at scale Sufficient compute to train and serve models at scale Dedicated support for publishing at top conferences Competitive SF salary and meaningful early equity Visa sponsorship and relocation support Generous health, dental, and vision coverage