Doyoung Kim
@doyoungkim-ml.bsky.social
Incoming CS PhD @NYU_Courant; Prieviously MS & BS @kaist_ai; General intelligence in Language ∪ Robotics
https://doyoungkim-ml.github.io/
https://doyoungkim-ml.github.io/
paper: arxiv.org/abs/2406.15275
How language models extrapolate outside the training data: A case study in Textualized Gridworld
Language models' ability to extrapolate learned behaviors to novel, more complex environments beyond their training scope is highly unknown. This study introduces a path planning task in a textualized...
arxiv.org
December 9, 2024 at 3:50 PM
paper: arxiv.org/abs/2406.15275
We characterize this CoT approach as 'Cognitive Maps for Path Planning' since it mirrors key aspects of human cognition, including extrapolation and rapid adaptation. We expect the development of a generalized cognitive map for language models to be a promising next step
December 9, 2024 at 3:50 PM
We characterize this CoT approach as 'Cognitive Maps for Path Planning' since it mirrors key aspects of human cognition, including extrapolation and rapid adaptation. We expect the development of a generalized cognitive map for language models to be a promising next step
Tl;dr:
Standard planning methods, whether implicit or with conventional CoT, fail to extrapolate beyond training data. However, we found a CoT variant configured with sampling, propagation, and backtracking, enabling robust planning with world model in novel environments.
Standard planning methods, whether implicit or with conventional CoT, fail to extrapolate beyond training data. However, we found a CoT variant configured with sampling, propagation, and backtracking, enabling robust planning with world model in novel environments.
December 9, 2024 at 3:50 PM
Tl;dr:
Standard planning methods, whether implicit or with conventional CoT, fail to extrapolate beyond training data. However, we found a CoT variant configured with sampling, propagation, and backtracking, enabling robust planning with world model in novel environments.
Standard planning methods, whether implicit or with conventional CoT, fail to extrapolate beyond training data. However, we found a CoT variant configured with sampling, propagation, and backtracking, enabling robust planning with world model in novel environments.
Especially fascinated by how humans solve complex problems through exploration and self-refinement - particularly interested in extrapolable learning (test time scale such as MCTS inference, in-context RL agent) and VLA agent for robotics.
Also going NeurIPS 2024! DM me for research chat or coffee
Also going NeurIPS 2024! DM me for research chat or coffee
November 24, 2024 at 7:41 PM
Especially fascinated by how humans solve complex problems through exploration and self-refinement - particularly interested in extrapolable learning (test time scale such as MCTS inference, in-context RL agent) and VLA agent for robotics.
Also going NeurIPS 2024! DM me for research chat or coffee
Also going NeurIPS 2024! DM me for research chat or coffee
Hi, can you add me?
November 24, 2024 at 5:51 PM
Hi, can you add me?
Hi, can you add me?
November 18, 2024 at 3:30 PM
Hi, can you add me?