# Can Large Language Models Infer Causation from Correlation?

**Authors:** Zhijing Jin, Jiarui Liu, Zhiheng Lyu, Spencer Poff, Mrinmaya Sachan, Rada Mihalcea, Mona Diab, Bernhard Schölkopf  
**arXiv:** [2306.05836](https://arxiv.org/abs/2306.05836)  
**Date:** 2023-06-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.05836v1/

## Key points


## Summary

Causal inference is a fundamental aspect of human intelligence and the field of CausalNLP has gained significant interest in recent years. Existing causal inference datasets in NLP primarily rely on discovering causality from empirical knowledge such as commonsense knowledge. To address this gap, a team of researchers proposed Corr2Cause, the first benchmark dataset to test the pure causal inference skills of large language models (LLMs). The task involves taking a set of correlational statements and determining the causal relationship between variables. The team curated a large-scale dataset comprising over 400K samples and evaluated seventeen existing LLMs on it. Their experiments identified a key shortcoming of LLMs in terms of their causal inference skills; they achieved almost random performance on the task, indicating that they lack pure reasoning skills for causal inference. Although re-purposing LLMs via finetuning somewhat mitigated this shortcoming, these models still failed to generalize beyond in-distribution settings where variable names and textual expressions used in queries were similar to those in the training set. The Corr2Cause task presents a challenging problem for LLMs and highlights areas for future research aimed at improving their pure reasoning skills and generalizability. The dataset is available online along with code for further exploration by interested parties. Additionally, the researchers proposed several potential solutions to improve LLMs' performance on this task, including incorporating domain-specific information or developing novel architectures that explicitly model causality. Overall, this work provides valuable insights into the limitations of current LLMs regarding causal inference and highlights opportunities for future research aimed at enhancing their capabilities in this critical area.