Research
Who we are
We are a team of researchers and engineers formerly at Meta AI, Amazon AGI, and Google, with experience spanning LLM evaluation, fairness, alignment, and embodied agents.
We study how models reason, plan, remember, and behave across long-horizon tasks. Our research connects these investigations with world modeling, synthetic data generation, and methods for task design and difficulty calibration.
Research Work
Research Focus Areas
We investigate failures such as sycophancy, misrepresentation, and violations of user intent. Our research develops methods to assess and improve how agents follow constraints, communicate uncertainty, and exercise judgment.
We study how agents reason, retain information, and track changing state across extended workflows. Our work examines failures in planning, memory, and collaboration across tools and applications.
We research world modeling, synthetic task generation, and reward design to scale useful training experiences. We study how environment realism, task diversity, and difficulty influence what models learn and how well they generalize.
.png)
