CLadder: Assessing Causal Reasoning in Language Models
About
The ability to perform causal reasoning is widely considered a core feature of intelligence. In this work, we investigate whether large language models (LLMs) can coherently reason about causality. Much of the existing work in natural language processing (NLP) focuses on evaluating commonsense causal reasoning in LLMs, thus failing to assess whether a model can perform causal inference in accordance with a set of well-defined formal rules. To address this, we propose a new NLP task, causal inference in natural language, inspired by the "causal inference engine" postulated by Judea Pearl et al. We compose a large dataset, CLadder, with 10K samples: based on a collection of causal graphs and queries (associational, interventional, and counterfactual), we obtain symbolic questions and ground-truth answers, through an oracle causal inference engine. These are then translated into natural language. We evaluate multiple LLMs on our dataset, and we introduce and evaluate a bespoke chain-of-thought prompting strategy, CausalCoT. We show that our task is highly challenging for LLMs, and we conduct an in-depth analysis to gain deeper insights into the causal reasoning abilities of LLMs. Our data is open-sourced at https://huggingface.co/datasets/causalNLP/cladder, and our code can be found at https://github.com/causalNLP/cladder.
Key insights
What problem does CLadder address?
It tests whether LLMs can perform formal causal reasoning rather than just recalling commonsense causal facts. The authors build a 10K-sample dataset of symbolic causal queries with ground-truth answers from a causal inference engine, verbalized into natural language.
How does this differ from prior work?
Unlike prior NLP work on commonsense causality, CLadder grounds questions in formal causal inference across all three rungs of causation. On the 10K benchmark, their CausalCoT prompting reaches 70.40% accuracy, improving vanilla GPT-4 by 8.37 points.
What are the practical uses of these findings?
CLadder serves as a benchmark for evaluating and improving LLM causal reasoning. The authors note reliable causal inference matters for policy-relevant scenarios in epidemiology, economics, and fairness, but document no actual deployment.
What limitations do the authors note?
The dataset covers only some commonly studied causal queries with binary variables and three-to-four-node graphs, not all possible combinations. The authors note this is a first step toward a mini-Turing test and suggest extending to further query types.