World models generally refer to generative models of the transition functions in an MDP, able to aid an agent in selecting an optimal set of actions to complete a given task. LLMs, for all their internet scale knowledge, still struggle to be effective world models (see here) for embodied agents. And yet there is incredible potential in using large-scale data especially in the form of language to aid agents, we as people are able to read information via language and form generalizable world models internally that helps us achieve tasks. This project focuses on realizing such potential.
- Do Embodied Agents Dream of Pixelated Sheep: Embodied Decision Making using Language Guided World ModellingIn International Conference on Machine Learning (ICML), 2023
- SwiftSage: A Generative Agent with Fast and Slow Thinking for Complex Interactive TasksIn Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS), 2023
- ScienceWorld: Is your Agent Smarter than a 5th Grader?In Empirical Methods in Natural Language Processing (EMNLP), 2022
- Modeling Worlds in TextIn The First Workshop on Commonsense Reasoning and Knowledge Bases (CSKB) at AKBC, 2021
- Modeling Worlds in TextIn Thirty-fifth Conference on Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track (Round 1), 2021
- Learning Knowledge Graph-based World Models of Textual EnvironmentsIn Thirty-fifth Conference on Neural Information Processing Systems (NeurIPS), 2021