Research Scientist (post-training)

San Francisco HQ • FullTime

Posted 19d ago

Job Location

San Francisco HQ

Tech Stack

Fine-Tuning PyTorch Reinforcement Learning RLHF Distributed Training LLMs diffusion models Evaluation Frameworks Generative Models AI Alignment video generation

Remote Work Policy

On-site

Employment Type

FullTime

About the job

Genmo is a research lab developing sophisticated video world models to understand, simulate, and interact with the physical world, aiming to unlock the right brain of AGI. We are seeking an exceptional Research Scientist to focus on alignment and post-training techniques for large-scale video generation models. This role is at the forefront of ensuring our diffusion-based video models reliably produce high-quality, physically accurate, and safe outputs that align with human preferences and values.

Responsibilities

Lead research in alignment and post-training methods for video generation models, focusing on quality, reliability, and human intent.
Design and implement supervised fine-tuning and RLHF pipelines for video generation models.
Develop robust evaluation frameworks to measure model alignment, safety, and output quality.
Create and optimize data collection pipelines for human feedback and preferences.
Design and conduct experiments to validate alignment techniques and their scaling properties.
Collaborate with cross-functional teams to integrate alignment improvements into production pipelines.
Review academic literature in generative AI and alignment to stay at the cutting edge.
Mentor junior researchers and foster responsible AI development.
Work closely with product teams to ensure alignment methods enhance model capabilities.

Requirements

Ph.D. in Computer Science, Artificial Intelligence, Machine Learning, or a closely related field.
Strong publication record in top-tier conferences (e.g., NeurIPS, ICML, ICLR) with a focus on reinforcement learning, alignment, or generative models.
Extensive experience implementing and optimizing large-scale training pipelines using PyTorch.
Deep understanding of reinforcement learning techniques, particularly RLHF.
Experience with distributed training systems and large-scale experiments.
Proven track record in designing and implementing robust evaluation frameworks.
Excellent communication skills to explain complex technical concepts.
Strong software engineering skills and experience with complex shared codebases.
Experience with diffusion models or other generative architectures (ideal).
Background in fine-tuning large language models or generative models (ideal).
Experience working with human feedback data collection and annotation pipelines (ideal).
Strong aesthetic sense and understanding of video quality assessment (ideal).
Familiarity with alignment techniques such as constitutional AI or debate (ideal).
Track record of successful collaboration with product teams (ideal).
Experience with perceptual quality metrics and human evaluation design (ideal).
Contributions to open-source projects in AI alignment or generative AI (ideal).

About Genmo

View company profile