Cognition
Research Engineer, Mid-Training
San Francisco · On-site · Full-time · Posted Aug 6, 2026
- Education
- A PhD is one signal among many; demonstrated capability is valued more than credentials.
About the job
Own late-stage LLM training decisions that improve model capabilities in coding, mathematics, reasoning, and long-context tasks, from data and schedules through evaluation and deployment.
What you'll do
- Design data mixtures and synthetic data pipelines for late-stage training, and improve data sourcing, filtering, and weighting.
- Research training interventions, annealing schedules, and context-length extension methods; measure their effects at scale.
- Build evaluations to verify capability gains and translate research into improvements for deployed agents.
What they ask for
- Deep familiarity with the end-to-end LLM training pipeline, including pre-training data, optimization, architecture, and interactions between mid-training and post-training.
- Hands-on experience with continual pre-training, annealing, or late-stage data mixing for large models.
- Proficiency in Python and PyTorch, including debugging distributed training at scale.
- Strong fundamentals in optimization, statistics, and ML theory.
Nice to have
- Experience developing or evaluating synthetic data pipelines for capability improvement.
- Original contributions demonstrated through publications, open-source impact, or internal results.
Tools
Python, PyTorch
AI focus
Mid-training and late-stage training research for LLMs, including data mixture design, synthetic data, capability improvements, annealing, context-length extension, and evaluation.
Model-written summary · the original posting is authoritative