This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
Modeling Open-World Cognition as On-Demand Synthesis of Probabilistic Models
Created by
Haebom
Author
Lionel Wong, Katherine M. Collins, Lance Ying, Cedegao E. Zhang, Adrian Weller, Tobias Gerstenberg, Timothy O'Donnell, Alexander K. Lew, Jacob D. Andreas, Joshua B. Tenenbaum, Tyler Brooke-Wilson
Outline
This paper explores the ability of people to make inferences and predictions by leveraging relevant information from diverse background knowledge when faced with a novel situation. We hypothesize that people construct sophisticated mental models tailored to new situations by combining distributed and symbolic representations, and propose a computational model, the “Model Synthesis Architecture (MSA),” that implements this idea. MSA performs global relevance-based retrieval and model synthesis using language models, and uses probabilistic programming to construct customized and consistent world models. We evaluate MSA using a novel inference dataset (based on modeled Olympic sports vignettes), and find that MSA better captures human judgment than a baseline model using only language models. This suggests that MSA can be implemented in a way that mirrors the human ability to make locally consistent inferences about globally relevant variables, and provides a way to understand and replicate human inference in open domains.
Takeaways, Limitations
•
Takeaways:
◦
Provides insight into the human reasoning process, which constructs sophisticated mental models for new situations through a combination of distributed and symbolic representations.
◦
We demonstrate that MSA, which combines language models and probabilistic programming, can effectively mimic human open-ended reasoning ability.
◦
MSA presents a novel approach to achieve human-level reasoning and contributes to the development of artificial intelligence in open fields.
•
Limitations:
◦
Currently, the performance of MSA is limited to a specific dataset (Model Olympics), and further research is needed on its generalizability to other domains.
◦
The complexity of MSA requires review of real-world applications and scalability.
◦
It cannot perfectly mimic the human reasoning process and may not encompass all aspects of human reasoning.
