Integrating Causal Inference with Large Language Models

LLMs enhance causal inference by leveraging their generative capabilities for tasks like treatment effect estimation and causal relationship discovery*, including generating counterfactuals and identifying latent variable interactions.
Applying causal frameworks to LLMs improves reasoning, commonsense understanding, bias mitigation, explainability, and safety* by identifying and correcting spurious correlations through causal interventions, structural modeling, and prompt manipulation.
* Causal reasoning ensures balanced use of textual and visual data in multimodal settings.
* Benchmarks are being developed to evaluate causal capabilities across various tasks, with hybrid approaches combining LLM outputs and traditional causal methods to refine causal discovery.
According to additional sources, causal inference evaluates and improves LLMs in reasoning capacity, fairness and safety, explainability, and handling multimodality*.
* Additional sources note that LLMs can directly recite causal knowledge without understanding true causality and perform worse than fine-tuned BERT in determining causality/correlation.