The paper "LLM Processes: Numerical Predictive Distributions Conditioned on Natural Language" by James Requeima explores the integration of prior knowledge and beliefs into predictive models using Large Language Models (LLMs). The authors aim to develop a regression model that can process numerical data and make probabilistic predictions guided by natural language text, which describes the user's prior knowledge. The key contributions include: 1. **Definition of LLM Processes (LLMPs)**: The paper introduces LLMPs, which are stochastic processes that can condition on both numerical data and unstructured text to improve predictive performance. LLMPs are defined over arbitrary-sized target sets and can handle multi-dimensional regression and density estimation. 2. **Prompt Engineering**: The authors investigate effective prompting practices for eliciting joint numerical predictions from LLMs. They explore various methods for conditioning LLMs on numerical data, including prompt formatting, ordering, and scaling. The results show that ordering training points by distance to the target point significantly improves performance. 3. **Performance Evaluation**: Extensive experiments demonstrate that LLMPs are competitive and flexible regressors, even on messy data. They achieve well-calibrated uncertainty and perform competitively with Gaussian Processes (GPs), LLMTIME, and Optuna. LLMPs can handle missing data, perform image reconstruction, and output multimodal predictive distributions. 4. **Incorporating Text into Predictions**: The paper shows that incorporating problem-relevant information through unstructured text into numerical predictions improves predictive performance. This allows users to provide prior, potentially expert, information in plain language and access problem-relevant latent knowledge encoded in LLMs. 5. **Related Work**: The authors discuss related work in LLM forecasting, regression, and in-context learning, highlighting how their approach differs from existing methods. 6. **Discussion, Limitations, and Societal Impact**: The paper discusses the potential societal impact of LLMPs, including the ability to generate probabilistic predictions using plain language, and acknowledges limitations such as computational costs and the need for more powerful LLMs. Overall, the paper advances the field of probabilistic modeling by demonstrating the feasibility of using LLMs to incorporate prior knowledge and text-based information into predictive models, making complex analyses more accessible to non-experts.The paper "LLM Processes: Numerical Predictive Distributions Conditioned on Natural Language" by James Requeima explores the integration of prior knowledge and beliefs into predictive models using Large Language Models (LLMs). The authors aim to develop a regression model that can process numerical data and make probabilistic predictions guided by natural language text, which describes the user's prior knowledge. The key contributions include: 1. **Definition of LLM Processes (LLMPs)**: The paper introduces LLMPs, which are stochastic processes that can condition on both numerical data and unstructured text to improve predictive performance. LLMPs are defined over arbitrary-sized target sets and can handle multi-dimensional regression and density estimation. 2. **Prompt Engineering**: The authors investigate effective prompting practices for eliciting joint numerical predictions from LLMs. They explore various methods for conditioning LLMs on numerical data, including prompt formatting, ordering, and scaling. The results show that ordering training points by distance to the target point significantly improves performance. 3. **Performance Evaluation**: Extensive experiments demonstrate that LLMPs are competitive and flexible regressors, even on messy data. They achieve well-calibrated uncertainty and perform competitively with Gaussian Processes (GPs), LLMTIME, and Optuna. LLMPs can handle missing data, perform image reconstruction, and output multimodal predictive distributions. 4. **Incorporating Text into Predictions**: The paper shows that incorporating problem-relevant information through unstructured text into numerical predictions improves predictive performance. This allows users to provide prior, potentially expert, information in plain language and access problem-relevant latent knowledge encoded in LLMs. 5. **Related Work**: The authors discuss related work in LLM forecasting, regression, and in-context learning, highlighting how their approach differs from existing methods. 6. **Discussion, Limitations, and Societal Impact**: The paper discusses the potential societal impact of LLMPs, including the ability to generate probabilistic predictions using plain language, and acknowledges limitations such as computational costs and the need for more powerful LLMs. Overall, the paper advances the field of probabilistic modeling by demonstrating the feasibility of using LLMs to incorporate prior knowledge and text-based information into predictive models, making complex analyses more accessible to non-experts.