Synthetic Data Generation and User Simulation PhD Research Intern — Fall 2026

What you'll do

• Research innovative techniques for high-fidelity synthetic data creation for LLMs.
• Calibrate simulated users against real behavioral signatures to enhance training signals.
• Develop and apply methods for trajectory generation guided by verification processes.
• Conduct experiments to validate that synthetic data improves downstream model performance.
• Extract process-reward models from multi-step interactions for better training outcomes.
• Collaborate on integrating novel methods into production pipelines for LLM training.

What we're looking for

• PhD candidate in Computer Science, Machine Learning, Computational Linguistics, or related field with deep learning specialization.
• Research experience in generative modeling, synthetic data generation, LLM post-training, reward modeling, and interactive simulation.
• Strong Python programming skills and hands-on experience with PyTorch and modern LLM training frameworks like HuggingFace.
• Experience training and evaluating LLMs on real downstream tasks, including calibration of LLM-as-judge for subjective dimensions.
• Prior work on user simulation grounded in real population data or cognitive science, focusing on behavioral modeling and agent–user interaction.
• Contributions to open-source projects related to SDG, LLM training, or evaluation.