A novel framework of training a contextualization module to help the decision-making of LLM agents achieves the super-human performance in the WebShop benchmark.

We propose a new benchmark for evaluating the safety and helpfulness of agents, with extensive analysis of the shortcomings of frontier LLM agents in mobile device control.

A novel benchmark that can serve as a unified testbed for mobile device control agents on performing practical daily tasks across diverse device configurations.

Reinforcement learning agents discover semantically meaningful skills with tasks proposed by a large language model and rewards from a vision-language model.

A newly proposed distribution (over a Riemannian manifold of the diagonal Gaussians equipped with Fisher information metric) empowers learning a hyperbolic world model.