Pika Research Scientist at Pika specializing in RL post-training and generative modeling for large-scale video generation. Responsible for running RL post-training, building video reward models, and leading post-training evaluation.
Responsibilities
At Pika, we are pioneering the next generation of creative infrastructure built around real-time, multimodal generation and intelligent agentic platforms. We are seeking Research Scientists with expertise in RL post-training and generative modeling for large-scale video generation. The focus is on refining Pika's video generation models using RL alignment and building robust video reward models. This is a staff and lead-level opportunity.
As a key member of our research team, you will own RL-based post-training for video diffusion/flow-matching models, develop state-of-the-art reward models, and lead post-training evaluation across human and automated metrics. You will collaborate closely with engineering and product teams, shaping the frontier of real-time creative and agentic video platforms.
RL alignment of Pika's video generation models and the reward models that drive them.
Distillation of RL-tuned models is a secondary focus.
Run RL post-training (preference optimization, online RL against learned rewards) for video diffusion/flow-matching models at multi-node scale.
Build video reward models: define target evaluation dimensions, design/configure preference data collection workflows, train and validate learned judges, and safeguard against reward hacking.
Own post-training evaluation, including human preference studies and their correlation with automated metrics.
Distill RL-tuned models to efficient few-step samplers while preserving alignment gains (secondary focus).
Qualification
Required
2+ years hands-on research experience in post-training or generative modeling.
RL or preference-optimization experience on generative models with evidence of model improvement.
Strong grounding in diffusion or flow-matching models, PyTorch, and multi-node distributed training.
Preferred
Experience developing reward models for visual generation, including VLM-as-judge or large-scale preference data collection.
Distillation expertise (distribution matching, consistency, adversarial approaches), ideally for video models.
Familiarity with video-specific failure modes: temporal drift, motion and physics realism.
