About the Role
As a Research Scientist working on Safety Post-Training at Scale Labs, you will develop and apply post-training methods and interpretability techniques to make frontier AI systems safer and better understood by researchers and policymakers. Scale plays an integral role in understanding the capabilities and safeguarding AI models and systems as a leading data and evaluation partner for frontier AI companies. Scale Labs has launched a new team focused on policy research, bridging the gap between AI research and global policymakers to enable informed, scientific decisions about AI risks and capabilities.
Our research addresses challenging problems in agent robustness, AI control protocols, and AI risk evaluations to help governments, industry, and the public understand and mitigate AI risk while maximizing AI adoption. This team collaborates broadly across industry, the public sector, and academia, regularly publishing findings.
Responsibilities
- Design and run post-training pipelines to study how training choices affect model safety, robustness, and alignment properties.
- Develop interpretability-informed evaluations that reveal how and why models produce unsafe, deceptive, or otherwise undesirable behaviors, and use those insights to guide targeted mitigations.
- Collaborate with policymakers, engineers, and other researchers to translate post-training and interpretability findings into actionable safety standards, evaluation benchmarks, and best practices.
Requirements
Ideally you’d have:
- Commitment to our mission of promoting safe, secure, and trustworthy AI deployments as frontier AI capabilities continue to advance.
- Experience with post-training and RL techniques such as RLHF, DPO, GRPO, and similar approaches.
- A track record of published research in machine learning, particularly in generative AI.
- At least three years of experience addressing sophisticated ML problems, whether in a research setting or in product development.
- Strong written and verbal communication skills to operate in a cross-functional team.
Nice to have:
- Experience with mechanistic interpretability, probing, or other techniques for understanding model internals.
- Familiarity with red-teaming or adversarial evaluation of post-trained models.
- Experience studying failure modes introduced or masked by post-training, such as reward hacking, sycophancy, or alignment faking.
Our research interviews are crafted to assess candidates' skills in practical ML prototyping and debugging, their grasp of research concepts, and their alignment with our organizational culture. We will not ask any LeetCode-style questions. If you’re excited about advancing AI safety and contributing to our mission, we encourage you to apply, even if your experience doesn’t perfectly align with every requirement.