LLMs can be used to generate numerical predictive distributions conditioned on natural language text, enabling probabilistic predictions at arbitrary locations. This approach allows users to incorporate prior knowledge into models through natural language, making them more accessible to non-experts. The paper introduces LLM Processes (LLMPs), which extend traditional probabilistic models to handle multi-dimensional regression and density estimation. LLMPs are evaluated on various tasks, including forecasting, image reconstruction, and black-box optimization, showing competitive performance with Gaussian Processes (GPs) and other models. The method uses prompt engineering to elicit joint predictive distributions from LLMs, with strategies for formatting, ordering, and scaling inputs. The paper also demonstrates the ability of LLMPs to incorporate textual information, improving predictive performance and providing quantitative structure that reflects qualitative descriptions. The results show that LLMPs can handle missing data, perform image reconstruction, and output multimodal predictive distributions. The work highlights the potential of LLMs to bridge the gap between expert knowledge and probabilistic modeling, enabling more flexible and interpretable models.LLMs can be used to generate numerical predictive distributions conditioned on natural language text, enabling probabilistic predictions at arbitrary locations. This approach allows users to incorporate prior knowledge into models through natural language, making them more accessible to non-experts. The paper introduces LLM Processes (LLMPs), which extend traditional probabilistic models to handle multi-dimensional regression and density estimation. LLMPs are evaluated on various tasks, including forecasting, image reconstruction, and black-box optimization, showing competitive performance with Gaussian Processes (GPs) and other models. The method uses prompt engineering to elicit joint predictive distributions from LLMs, with strategies for formatting, ordering, and scaling inputs. The paper also demonstrates the ability of LLMPs to incorporate textual information, improving predictive performance and providing quantitative structure that reflects qualitative descriptions. The results show that LLMPs can handle missing data, perform image reconstruction, and output multimodal predictive distributions. The work highlights the potential of LLMs to bridge the gap between expert knowledge and probabilistic modeling, enabling more flexible and interpretable models.