The opportunity
Cohere is the leading security-first enterprise AI company. We build cutting-edge foundation AI models and end-to-end products that are designed to solve real-world business problems.
What you'll do
Design, run, and analyse large-scale RL post-training experiments (e.g., RLHF, RL from verifiable rewards).
Research and develop methods that improve out-of-distribution generalization,: such as reward design, environment and task diversity, curricula, and regularization.
Build held-out evaluations that measure generalization to unseen tasks,: domains, and formats, and that catch reward hacking and benchmark overfitting.
Diagnose training dynamics and failure modes, including instability, reward: over-optimization, and capability regressions.
Turn successful research into robust, reproducible training recipes for production models.
Strong research and software engineering skills, with hands-on experience training LLMs.
What they're looking for
- Experience applying reinforcement learning to LLMs (e.g., PPO, GRPO, RLHF, RLVR), ideally at scale.
- Track record of work on generalization, robustness, or distribution shift,: through publications or shipped models.
- Rigorous experimental practice: clean ablations, careful controls, and sound conclusions from noisy results.
- Experience with distributed training (e.g., JAX, PyTorch, FSDP, Megatron) is a plus.
