Researcher, Multimodal Safety

Overview

Researcher for the Chat and Multimodal Safety team, working on multimodal safety research and methods to make frontier models respond safely across text, vision, and audio.

What you'll do

• Define and advance multimodal safety research linking perception and semantic understanding to safe model behavior.
• Build training and evaluation methods for vision-language models (VLMs) including post-training, safety evaluations, and interventions.
• Develop approaches for models to respond safely and appropriately in varied contexts.
• Collaborate with Personal AGI, Consumer Devices, and product/model teams to translate research into safer ambient, embedded, and personalized experiences.

What you'll need

• Track record building or advancing multimodal models across vision-language, video understanding, image generation, audio, or multimodal reasoning.
• Fluency across perception and language.
• Ability to understand multimodal systems end to end from encoders and modality fusion to cross-modal reasoning and inference tradeoffs.
• Experience improving frontier model behavior through post-training (e.g., SFT, RL, data curation, synthetic data, evaluation, and error analysis).
• Strong research and engineering judgment for open-ended safety problems, including testable hypotheses and decisive experiment/evaluation design.

Details

• Role based in San Francisco, CA.
• Hybrid work model with 3 days in the office per week.
• Relocation assistance offered to new employees.