ML Researcher - Posttraining

Role overview

Collaborate with our AI research and engineering teams to integrate advancements into our products. We're looking for someone with deep experience in large-scale finetuning of diffusion models and posttraining techniques. What you'll doFinetune diffusion models at scale to improve image aesthetics and quality. Design comprehensive eval suites and reward designs for the reinforcement learning stage focused on image space.

What you'll do

• Finetune diffusion models at scale to improve image aesthetics and quality.
• Implement posttraining techniques ranging from supervised finetuning, preference optimization, reinforcement learning, on-policy distillation, and various distillation / acceleration techniques.
• Design comprehensive eval suites and reward designs for the reinforcement learning stage focused on image space.
• Train custom VLM as reward models as part of our reward design.
• Train custom LLMs for prompt expansion through finetuning and reinforcement learning.
• Coordinate with data teams and partners to manage collection of preference data and model evaluation results.
• Work on safety alignment of our models for open source release.
• Collaborate with our AI research and engineering teams to integrate advancements into our products.

What the employer is looking for

• Proven work of posttraining diffusion models for image or video generation.
• Experience with large-scale model training, inference, and optimization.
• Strong understanding of both LLM and diffusion post training pipelines and algorithms such as PPO, GRPO, DPO, OPD, and MOPD.
• Strong proficiency in PyTorch and understanding of its inner workings.
• Strong background in distributed training paradigms such as FSDP, CP, SP, USP, TP, and EP. Knowing how different parallelism strategies work together and their tradeoffs.
• Good knowledge of low precision training / inference in FP8, NVFP4, and MXFP8.
• Good understanding of algorithms and techniques used in fast inference engines such as vLLM and sglang as well as existing RL frameworks in LLM space such as slime, miles, tinker, and verl.
• Understanding of various RL infrastructure and optimization techniques such as async RL, fast weight transfer, pipelining rollouts, managing off policy data.
• Ability to monitor model regression and identify weak areas and turn them into concrete evals and reward design.
• Experience training VLM models. Many of our custom reward models use VLM to provide reward signals for our models.
• Keeping up with the developments in related fields such as LLM, VLM, representation learning, and robotics research.
• Being comfortable working in a goal-oriented research environment.
• Having good judgement around when one should explore different training strategies and when it's time to commit to a specific strategy to scale compute and data.
• Comfortable working with underspecified goals. We expect every technical member to take an ambiguous research goal and break it down into concrete requirements, plans, experiment plan, and execution items.
• Good research taste — bias towards simplicity and methods that scale well with compute, data, and minimal human supervision.