Research Scientist, Reinforcement Learning (LLM) and Post-training

Advanced Micro Devices, Inc•Santa Clara, CA

4d•Hybrid

About The Position

We are hiring a Research Scientist, Reinforcement Learning (LLM) and Post-Training, specializing in reinforcement learning to advance post-training and interactive learning for large generative models applied to demanding engineering and hardware-adjacent tasks (code, optimization, tool use, and long-horizon decision making). You will invent and analyze RL algorithms—policy optimization, preference-based methods, exploration, credit assignment, and reward modeling—run rigorous empirical studies, and partner with infra and product teams to land methods that improve measurable task success without sacrificing stability or safety.

Requirements

You publish and ship.
You are fluent in both RL theory and the practical path from ablation to production-scale training.
You care about reward misspecification, variance reduction, and evaluation that reflects real constraints—not only toy environments.
PhD in Computer Science, Machine Learning, or related field strongly preferred.

Nice To Haves

Strong publication record in reinforcement learning or closely related machine learning areas.
Hands-on experience training RL or preference-optimized models at non-trivial scale (GPUs, distributed jobs)
Experience with LLM post-training, RLHF/RLAIF, or policy optimization for language or code agents
Familiarity with compilers, kernels, EDA-style workflows, or large-scale codebases is a plus

Responsibilities

Research and develop RL methods for post-training LLMs and code models on structured engineering tasks with verifiable or preference-based feedback
Design reward models, curricula, and off-policy or on-policy training recipes suited to sparse, noisy, or expensive labels from experts and simulators
Characterize failure modes (reward hacking, degenerate policies, instability) and propose mitigations grounded in experiments
Collaborate with RL infra engineers to scale training; define interfaces for rollout generation, logging, and reproducibility
Publish at top venues (e.g. NeurIPS, ICML, ICLR) and contribute internal technical leadership on the RL roadmap