Despite a model’s impressive ability to navigate, researchers found it did not form an accurate internal map, indicating a reliance on surface-level pattern recognition rather than true comprehension.
The study highlights that generative AI models may not learn coherent world models, which raises concerns about their reliability in real-world applications, especially under changing contexts.
Leading researcher Ashesh Rambachan noted the importance of understanding whether LLMs can represent reality accurately, especially if we want to leverage these models in scientific discoveries.
Researchers observed that when streets were closed, the AI's navigation performance deteriorated significantly, revealing that it could not adapt well to altered conditions.
Collection
[
|
...
]