LLM Research Scientist, Pre-training, Computer Vision, Adversarial Robustness

Role Overview

Advance well-scoped, open-ended empirical machine learning research across computer vision and language modeling. This role focuses on training, improving, evaluating, and deploying deep learning models under practical data, compute, model-size, robustness, and latency constraints.

Key Responsibilities

• Train image classifiers and generative image models from scratch, and fine-tune open-weight language models.
• Develop effective models within limited data, compute, and parameter budgets.
• Improve robustness against adversarial image inputs and adversarial conversational behavior.
• Compress models to satisfy strict size and latency requirements while maintaining accuracy.
• Investigate, diagnose, and resolve model-training issues.

Qualifications

• At least 3 years of machine learning research experience. PhD research counts toward this requirement.
• Strong experience with PyTorch, JAX, TensorFlow, or similar machine learning frameworks.
• A degree from a top-100 university, experience at a FAANG or comparable AI company, or an equivalent research record demonstrated through publications or impactful open-source contributions.
• Deep expertise in one or more of the following areas:
• Adversarial robustness, including PGD-based adversarial training, TRADES, robust-accuracy evaluation under standard threat models such as L∞ attacks and AutoAttack, avoiding gradient-masking issues, and managing robustness-accuracy trade-offs and robust overfitting.
• Efficient computer vision, including end-to-end image-classifier training, particularly fine-grained recognition with many visually similar classes and few examples per class; quantization, pruning, and knowledge distillation; and deployment in on-device, edge, or embedded environments with strict size or latency budgets.
• Generative image modeling, including training diffusion models, GANs, VAEs, or flow-based models from scratch; iterating against sample-quality measures such as FID; and applying training-efficiency techniques for high-quality, small-parameter generators.
• LLM post-training and behavioral robustness, including supervised fine-tuning; preference optimization through DPO, RLHF, or RLAIF; creating datasets through synthetic generation, noisy or weak supervision, and rejection sampling; multi-turn conversational behavior involving resistance to persuasion and sycophancy, calibrated confidence, and appropriate acceptance of corrections; and alignment-oriented fine-tuning that changes a specific behavior while preserving general capabilities.
• Multilingual pre-training, including training multilingual or low-resource-language models from scratch, tokenizer design across scripts and typologically diverse languages, and balancing unequal language-data volumes through sampling temperatures and cross-lingual transfer in data-constrained settings.
• Experience with scaling laws, training efficiency, curriculum learning, data ordering, benchmark construction, contamination control, statistically sound model comparisons, uncertainty estimation, model calibration, data augmentation, or synthetic data for robustness is a plus.

Work Terms

• Remote, hourly engagement.
• Flexible, project-based work focused on challenging, high-impact machine learning research in collaboration with leading AI researchers.

Compensation

• $100 to $120 per hour.