Research Scientist - RL Training

ABOUT THE ROLE 

We are looking for a Research Scientist to advance reinforcement learning for large language models. You will develop, implement, and run training methods that help models solve frontier-level challenging tasks, especially complex, multi-turn, long-horizon problems with different forms of feedback and interactions. By connecting RL research with Snorkel’s ability to build high-quality, domain-specific tasks, environments, and data—and using training and evaluation results to improve both—you will help push model capabilities forward and shape a distinctive research direction for Snorkel.

You will lead research and development on RL training for large language models. This includes designing algorithms and training recipes, implementing them in our scalable framework, and evaluating their impact across diverse tasks and environments. You will diagnose training behavior and model failures, then work with the data team to identify gaps in tasks, environments, or training signals and assess whether changes improve results. A key focus is extending RL to modern settings such as extremely complex, multi-turn, long-horizon tasks with various forms of feedback.

MAIN RESPONSIBILITIES 

• Research and implement cutting-edge reinforcement learning algorithms and training recipes for large language models, adapting and extending methods such as GRPO for complex, multi-turn, and long-horizon tasks; evaluate their effects on learning, generalization, and performance across diverse environments.
• Design and build data pipelines that generate high-quality training signals for RL workflows, including AI-assisted data annotation and curation data pipelines to improve model generalization to unseen benchmarks . 
• Prototype and iterate on end-to-end RL training recipes, using training and evaluation results to identify what improves performance and where tasks, environments, or training signals need to change.
• Work with research scientists, ML engineers, data teams, and delivery teams to diagnose training results, turn findings into actionable improvements to data and environments, and validate those improvements in subsequent training runs.
• Stay current with advances in large-scale distributed LLM training and scalable RL, and apply relevant methods to our training systems and challenging environments, including tool-use and other complex tasks.
• Contribute to Snorkel's research publications and internal knowledge base in RL and model training.

PREFERRED QUALIFICATIONS 

• Deep expertise in reinforcement learning for language models, including policy optimization, credit assignment, and learning from sparse or delayed rewards; experience designing or adapting methods for multi-turn and long-horizon tasks.
• Experience training or fine-tuning 30B+ large language models at scale, including familiarity with distributed training infrastructure. 
• Strong proficiency in Python and ML frameworks, especially PyTorch and HuggingFace and hands-on experience with RL frameworks such as Verl. 
• Strong software engineering skills, with a track record of building research systems and training pipelines that others can run, extend, and maintain.
• Familiarity with ML infrastructure and cloud platforms and tools (AWS, GCP, Kubernetes, Slurm, etc.); experience with large-scale RL training pipelines a strong plus. 
• Ability to analyze training behavior and evaluation results, identify likely causes of failures or regressions, and turn findings into concrete improvements to training methods, tasks, environments, or data.
• Ph.D. in machine learning, reinforcement learning, or a related field strongly preferred; exceptional industry experience considered.

Salary range(s) for this role 200000 - 375000 USD