Description
Today, NVIDIA is tapping into the unlimited potential of AI to define the next era of computing. An era in which our GPU acts as the brains of computers, robots, and self-driving cars that can understand the world.
As an NVIDIAN, you'll be immersed in a diverse, encouraging environment where everyone is inspired to do their best work. Come join the team and see how we can make a lasting impact on the world.
We're a research team dedicated to a major challenge in modern model development. It involves advanced artificial data creation across pre-training, post-training, and evaluation infrastructure. Collecting only real data at scale carries meaningful quality, cost, latency, and privacy tradeoffs; it tends to overrepresent certain populations; and it often leaves gaps on the long tail of languages, domains, demographics, and safety scenarios.
We're investigating how generative models can create instructional and assessment data that shows high utility. The measurement is based on downstream model performance instead of surface plausibility. Additionally, we explore grounding that data in real-world distributions to ensure it generalizes.
A major workstream within this agenda is population-grounded user simulation: synthetic users interacting with LLMs, calibrated against real behavioral signatures, and structured to yield training signals (SFT examples, preference pairs, verifier corpora, process reward models, on-policy RL environments).
Other examples include verifier-grounded trajectory synthesis where ground truth exists, multilingual and low-resource coverage, and SDG quality measurement across pre- and post-training corpora.
This is an opportunity to contribute to foundational research that will help shape how the next generation of AI models is trained.
Responsibilities:
- Researching innovative techniques in generative models, artificial data creation, user simulation, reward modeling, and data-quality estimation for LLM training.
- Crafting and applying new methods for high-fidelity synthetic data. For example, behavioral calibration of simulated users against real-user signatures. Also, procedurally generated probe and scenario coverage, trajectory generation guided by verification, process-reward extraction from multi-step interactions, and population-aware data mixing for pre-training and post-training.
- Conducting experiments to validate that your synthetic data measurably improves downstream model performance , accuracy, robustness, calibration, multilingual parity, agentic safety , rather than only matching surface statistics.
- Collaborating with other researchers and engineers to integrate novel methods into production training and evaluation pipelines.
- Preparing research findings for internal presentations and potential publication at top-tier AI conferences
Requirements:
- Pursuing a PhD in Computer Science, Machine Learning, Computational Linguistics, Computational Neuroscience, or equivalent program, with a specialization in deep learning, NLP, or LLM training.
- Research experience in at least one of: generative modeling, synthetic data generation, LLM post-training (SFT/RLHF/DPO/RL), reward modeling, multi-agent or interactive simulation, behavioral or cognitive modeling, or large-scale data curation.
- Excellent Python programming skills.
- Hands-on experience with deep learning frameworks (PyTorch) and the modern LLM training/serving stack (e.g., HuggingFace, vLLM, distributed training).
- Strong research background with publications at top-tier AI, ML, or NLP conferences
Benefits:
- Our internship hourly rates are a standard pay based on the position, your location, year in school, degree, and experience. The hourly rate for our interns is 30 USD - 94 USD.
- You will also be eligible for Internbenefits.