Joy Hsu Joins OpenAI for Agent Safety, Named Imperial College Assistant Professor

Joy Hsu Joins OpenAI and Imperial College

On November 4, Joy Hsu announced on X that she has joined OpenAI to work on improving AI agent safety in the real world. She will also serve as an assistant professor in the Department of Computing at Imperial College London.

Academic Background

Hsu completed her bachelor's, master's, and PhD in Computer Science at Stanford University under the supervision of Associate Professor Jiajun Wu. During her PhD, she was awarded the Knight-Hennessy Scholarship (a multidisciplinary program selecting up to 100 scholars globally each year, providing up to three years of funding for tuition, stipend, and academic expenses) and the NSF Graduate Research Fellowship.

Research Trajectory

2023: LEFT Framework for Concept Grounding (NeurIPS 2023)

As first author, Hsu published "What's Left? Concept Grounding with Logic-Enhanced Foundation Models" at NeurIPS 2023. The LEFT framework combines logic programs generated by large language models with trainable domain-specific modules, enabling AI to learn and understand abstract concepts across 2D images, 3D scenes, human motion, and robotic manipulation, and transfer reasoning to unseen complex tasks.

2024: PARL Framework for Planning Abstractions from Language (ICLR 2024)

Hsu contributed to "Learning Planning Abstractions from Language" , accepted at ICLR 2024. The PARL framework allows AI to learn transferable abstract planning capabilities from language demonstrations, adapting to new environments, unfamiliar object combinations, and longer-horizon tasks.

2025: DSG Framework for Visual Abstraction Reasoning (ICLR 2025)

Hsu participated in "What Makes a Maze Look Like a Maze?" , accepted at ICLR 2025. The research investigates how AI understands implicit abstract concepts in images—for example, recognizing that a specific arrangement of branches forms maze walls. The DSG framework uses an LLM to extract a dependency graph of abstract concepts, then combines a vision-language model to map those concepts layer by layer to concrete image elements, improving reasoning about complex visual abstractions.

2026: SAW-BENCH for Situated Awareness in the Real World (ICML 2026 Spotlight)

The paper "SAW-BENCH: Learning Situated Awareness in the Real World" was accepted at ICML 2026 with a Spotlight distinction. It introduces a benchmark for real-world situated awareness using first-person video captured by smart glasses to evaluate multimodal models' understanding of their own location, motion state, and surrounding spatial relations. The benchmark reveals a significant gap between current models and human-level spatial perception in real-world settings.

Transition to Agent Safety

Across cross-domain concept learning, visual abstraction reasoning, and real-world situated awareness, Hsu's work consistently centers on enabling AI to better understand environments, build spatial cognition, and perform reliable reasoning. Her move to OpenAI and Imperial College marks a new phase combining academic research with frontier industry exploration focused on AI agent safety.