Cognition

Research role focused on post-training and alignment of AI agents, shaping their behavior and capabilities for real-world usefulness and safety. Blends research and engineering to iterate on datasets, training stages, hyperparameters, and evaluation design, with a focus on understanding and advancing techniques like RLHF/RLAIF for long-horizon tasks.

This role focuses on the critical mid-training stage of LLMs, bridging pre-training and post-training. The goal is to sharpen raw base model capabilities in areas like reasoning, generalization, coding, and math through strategic data mix, quality uplift, annealing schedules, context length extension, and synthetic data generation. The role involves both research and engineering, with a strong emphasis on evaluation and iteration to ensure measurable capability gains for AI agents like Devin.