Research Scientist, Safety Post Training

Written by the T&S Network from Scale AI's posting.

Post-training and interpretability research on frontier AI models, aimed at making model safety behavior understood by researchers and policymakers. The work sits in Scale Labs' policy research team, which bridges AI research and global policymakers on AI risk and capabilities.

• AI misuse and jailbreaks

What you'd do

• Build and run post-training pipelines that test how training decisions shape safety, robustness, and alignment.
• Build interpretability-informed evaluations that show how and why models produce unsafe, deceptive, or undesirable behavior.
• Use those findings to guide targeted mitigations for unsafe model behavior.
• Turn post-training and interpretability results into safety standards, benchmarks, and best practices.
• Work with policymakers, engineers, and researchers on a cross-functional team.

What it asks for

• Three or more years of hands-on work on hard ML problems in research or product settings.
• Hands-on work with post-training and RL methods like RLHF, DPO, and GRPO.
• Publication record in machine learning, especially generative AI.
• Clear written and spoken communication for cross-functional work.
• Work with mechanistic interpretability, probing, or other ways of studying model internals (preferred).
• Exposure to red-teaming or adversarial testing of post-trained models (preferred).
• Experience studying failure modes such as reward hacking, sycophancy, or alignment faking (preferred).

Worth knowing

• Full-time position.

Terms in this role

Red teaming.

Testing a system, often an AI model, by deliberately trying to make it fail or misbehave, so problems are found and fixed first.

RLHF.

Reinforcement learning from human feedback: training an AI model on people's ratings of its answers.

Our best summary of the posting, and it can be wrong. The employer's posting is what counts: read it in full on Scale AInorth_east before you apply. If we got something wrong, report this listing.