Job Description

We are looking for a research scientist to push the frontier of multimodal representation learning for autonomous driving. You will focus on method-level innovation, including designing new learning paradigms, architectures, and training strategies for large-scale vision and video models.

This role is ideal for candidates who care deeply about how models learn, not just how to train them.

Responsibilities

1. Design and develop novel representation learning methods, including:

• Self-supervised learning
• Contrastive learning
• Masked modeling
• Teacher-student / distillation frameworks

2. Innovate on model architectures, including:

• Vision Transformers (ViT)
• Video models (e.g., Video Swin)
• Multimodal fusion architectures

3. Develop new training paradigms and loss functions to improve representation quality

4. Drive measurable improvements in:

• Representation quality (e.g., linear probe, retrieval metrics)
• Downstream performance (e.g., perception / E2E driving systems)

5. Collaborate with engineering and data teams to integrate new methods into production systems

Requirements

Core Requirements

1. Strong academic background in Computer Science, AI, or related fields
2. Solid understanding of modern representation learning frameworks:

• CLIP
• DINOv2
• Masked Autoencoders

3. Proven experience in method-level innovation, such as:

• Designing new loss functions
• Modifying architectures

4. Creating new training strategies

Preferred Qualifications

1. Publications in top-tier venues:

• CVPR
• ICCV
• NeurIPS
• ICLR

2. Experience training models at scale using:

• DeepSpeed / Megatron-LM

3. Strong intuition for why models behave the way they do, not just how to train them