LaMPP: Language Models as Probabilistic Priors for Perception and Action
This work addresses the challenge of handling uncertain observations in AI systems, though it appears incremental by applying existing language model techniques to new domains.
The paper tackles the problem of improving performance on non-linguistic perception and control tasks, such as semantic segmentation and household navigation, by using language models as probabilistic priors, resulting in enhanced predictions for rare and out-of-distribution inputs.
Language models trained on large text corpora encode rich distributional information about real-world environments and action sequences. This information plays a crucial role in current approaches to language processing tasks like question answering and instruction generation. We describe how to leverage language models for *non-linguistic* perception and control tasks. Our approach casts labeling and decision-making as inference in probabilistic graphical models in which language models parameterize prior distributions over labels, decisions and parameters, making it possible to integrate uncertain observations and incomplete background knowledge in a principled way. Applied to semantic segmentation, household navigation, and activity recognition tasks, this approach improves predictions on rare, out-of-distribution, and structurally novel inputs.