Member of Technical Staff, Model Training & Post-Training
Advance the quality, efficiency, and capabilities of world, video, image, and audio models. You will work across large-scale model training, post-training, efficient fine-tuning, and agentic systems that help creators plan, generate, evaluate, and refine multimodal content.
We are looking for exceptional researchers and engineers with depth in one or more of these areas—not necessarily all of them.
What you’ll do
- Train and improve world, video, image, and audio models through continued pre-training, mid-training, supervised fine-tuning, preference optimization, reinforcement learning, and other post-training methods
- Develop data recipes, objectives, reward signals, and evaluation frameworks that improve perceptual quality, controllability, temporal consistency, physical fidelity, and instruction following
- Adapt and optimize models using techniques such as distillation, quantization-aware training, pruning, sparsity-aware training, adapters, and low-rank fine-tuning
- Co-design training and inference methods to meet real-world latency, throughput, memory, and cost targets
- Post-train LLM and multimodal agents that can reason over creative intent, select and use tools, execute multi-step workflows, critique generated outputs, and iteratively improve image, video, and audio results
- Build training environments, graders, synthetic-data pipelines, and feedback loops for agentic multimodal workflows
- Design rigorous offline and online evaluations, analyze model and agent failure modes, and translate findings into decisive experiments
- Work closely with ML systems, performance, product, and customer-facing teams to bring research into production
- Strong experience with PyTorch, JAX, or a comparable deep-learning framework
- Hands-on experience training, fine-tuning, post-training, or optimizing large generative or multimodal models
- Depth in at least one of:
- World, video, image, or audio model training
- Post-training, reinforcement learning, reward modeling, or preference optimization
- Model distillation, quantization-aware training, pruning, sparsity, or efficient fine-tuning
- LLM or multimodal agents, tool use, or long-horizon workflow training
- Ability to turn ambiguous quality or product goals into datasets, training signals, experiments, and measurable evaluations
- Strong research and engineering judgment, including the ability to diagnose failed experiments and model behavior
- Comfort working onsite with the team in the Bay Area
Preferred qualifications
- Experience with modern generative architectures such as diffusion models, diffusion transformers, flow matching, autoregressive models, Transformers, or multimodal foundation models
- Experience running large-scale distributed training or post-training workloads
- Experience with SFT, DPO or related preference methods, RLHF/RLAIF, online RL, reward models, verifiers, or learned evaluators
- Experience optimizing generative models without materially degrading perceptual quality or controllability
- Experience building agent environments, tool-use datasets, graders, synthetic trajectories, or production agent harnesses
- Familiarity with creative workflows and tools for image, video, audio, 3D, or world-model applications
- A strong publication, open-source, or production track record in relevant areas