Large Language Models as Noisy Oracles for Constructing Causal Models
PI: Vasant Honavar (Informatics and Intelligent Systems)
PI will be responsible for tuition support.
Project Summary
Causal models (directed acyclic graphs that express causal knowledge) play a central role in representing and reasoning with causal assumptions, identifying confounders, determining whether a causal effect can be estimated from observational data, and if so, how, etc.). Typically, specifying causal models requires substantial domain knowledge which can be hard to come by. In principle, it is possible to rule out incorrect causal models if given access to an oracle that can answer arbitrary conditional independence queries. It is also possible to identify the correct causal model if given access to results of specific interventional experiments. Large language models (LLM) trained have been shown to be capable of extracting and encoding domain knowledge from literature. LLM can thought of as noisy oracles that can answer arbitrary natural language queries. This project aims to explore the use of LLMs as noisy oracles that can answer conditional independence queries or interventional queries to construct causal models.
The project has three main aims:
1. Identify specific classes of causal queries that can be reliably answered by LLMs.
2. Develop strategies for combining the answers to causal queries into a causal model
3. Assuming a certain error rate for the LLM-based noisy oracle for causal queries, establish a bound on the error of the (structure of the) causal model
4. Evaluate the resulting methods in synthetic settings (where the underlying causal structure is known and noisy causal oracle can be simulated) and in real-world settings where LLM is used as a noisy causal oracle.
Expected Impact
This project will yield theoretical foundations and practical methods with strong performance guarantees for constructing causal models in settings where domain knowledge is hard to come by. The resulting methods will find applications across a broad range of scientific domains (e.g., medicine, economics, social sciences) and contribute to advances in AI-enabled scientific discovery.
Junior Researcher Background
The junior researcher interested in working on this project should ideally possess strong background in causal modeling and causal inference (at a level comparable to that offered by DS 560 at Penn State) and working familiarity with state-of-the-art large language models, their capabilities, and limitations.