ConfigWM uses task language to filter distractors in world-model states
Original title: World Models with Configurable Perception
Authors report ConfigWM beats the strongest baseline by 22.3% on average as visual noise rises.
Researchers introduce ConfigWM, a world model that treats a language task description as a configuration signal for state extraction. Visual embeddings are mapped into a task-specific state space that keeps task-relevant detail and suppresses distractors, then action-conditioned dynamics are learned for planning. In noisier environments, existing methods lose planning success while ConfigWM stays comparatively stable. The authors also report higher physical-quantity probing correlation and better distracted-to-clean retrieval.
Unconfirmed: Performance figures are the authors' reported results, not independently verified.
- ConfigWM outperforms the strongest competing method by 22.3% on average.
- Physical-quantity probing rises by 0.365 Pearson r.
- Distracted-to-clean retrieval improves by 75.5%.
- A project page is at configwm.github.io.
From the original article
1 min read at huggingface.co
Real-world observations are noisy and often contain excessive task-irrelevant information. Effective world modeling and planning need to abstract observations into world state representations while filtering out irrelevant information. However, encoders in existing world models are typically non-adaptive, leading to state representations that are easily distracted by task-irrelevant information.
By Yijun Shen, Delong Chen, Xianming Hu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.