TL;DR
Research Scientist at Stability AI designing and fine-tuning large-scale Vision-Language Models for multimodal tasks including visual reasoning, 3D understanding, and embodied interaction. Bridges research breakthroughs with scalable training pipelines and production deployment.
Multimodal Generative AI Researcher
Location: Remote
About the Role
We’re looking for a Research Scientist with deep expertise in training and fine-tuning large Vision-Language and Language Models (VLMs / LLMs) for downstream multimodal tasks. You’ll help push the next frontier of models that reason across vision, language, and 3D, bridging research breakthroughs with scalable engineering.
What You’ll Do
- Design and fine-tune large-scale VLMs / LLMs — and hybrid architectures — for tasks such as visual reasoning, retrieval, 3D understanding, and embodied interaction.
- Build robust, efficient training and evaluation pipelines (data curation, distributed training, mixed precision, scalable fine-tuning).
- Conduct in-depth analysis of model performance: ablations, bias / robustness checks, and generalisation studies.
- Collaborate across research, engineering, and 3D / graphics teams to bring models from prototype to production.
- Publish impactful research and help establish best practices for multimodal model adaptation.
What You Bring
- PhD (or equivalent experience) in Machine Learning, Computer Vision, NLP, Robotics, or Computer Graphics.
- Proven track record in fine-tuning or training large-scale VLMs / LLMs for real-world downstream tasks.
- Strong engineering mindset — you can design, debug, and scale training systems end-to-end.
- Deep understanding of multimodal alignment and representation learning (vision–language fusion, CLIP-style pre-training, retrieval-augmented generation).
- Familiarity with recent trends, including video-language and long-context VLMs, spatio-temporal grounding, agentic multimodal reasoning, and Mixture-of-Experts (MoE) fine-tuning.
- Awareness of 3D-aware multimodal models — using NeRFs, Gaussian splatting, or differentiable renderers for grounded reasoning and 3D scene understanding.
- Hands-on experience with PyTorch / DeepSpeed / Ray and distributed or mixed-precision training.
- Excellent communication skills and a collaborative mindset.
Bonus / Preferred
- Experience integrating 3D and graphics pipelines into training workflows (e.g., mesh or point-cloud encoding, differentiable rendering, 3D VLMs).
- Research or implementation experience with vision-language-action models, world-model-style architectures, or multimodal agents that perceive and act.
- Familiarity with efficient adaptation methods — LoRA, adapters, QLoRA, parameter-efficient finetuning, and distillation for edge deployment.
- Knowledge of video and 4D generation trends, latent diffusion / rectified flow methods, or multimodal retrieval and reasoning pipelines.
- Background in GPU optimisation, quantisation, or model compression for real-time inference.
- Open-source or publication track record in top-tier ML / CV / NLP venues.
Equal Employment Opportunity:
We are an equal opportunity employer and do not discriminate on the basis of race, religion, national origin, gender, sexual orientation, age, veteran status, disability or other legally protected statuses.