[Paper Review] CausalBench: A Comprehensive Benchmark for Causal Learning Capability of LLMs

Yu Zhou, Xingyu Wu+4 more|arXiv (Cornell University)|Apr 9, 2024

Topic Modeling4 citations

TL;DR

CausalBench is a comprehensive benchmark evaluating large language models' (LLMs) causal learning capabilities across diverse tasks, causal network scales, and data types. It reveals that LLMs struggle with distinguishing correlation from causation, especially in complex or large-scale networks, and highlights the critical role of structured data and background knowledge in improving performance.

ABSTRACT

The ability to understand causality significantly impacts the competence of large language models (LLMs) in output explanation and counterfactual reasoning, as causality reveals the underlying data distribution. However, the lack of a comprehensive benchmark currently limits the evaluation of LLMs' causal learning capabilities. To fill this gap, this paper develops CausalBench based on data from the causal research community, enabling comparative evaluations of LLMs against traditional causal learning algorithms. To provide a comprehensive investigation, we offer three tasks of varying difficulties, including correlation, causal skeleton, and causality identification. Evaluations of 19 leading LLMs reveal that, while closed-source LLMs show potential for simple causal relationships, they significantly lag behind traditional algorithms on larger-scale networks ($>50$ nodes). Specifically, LLMs struggle with collider structures but excel at chain structures, especially at long-chain causality analogous to Chains-of-Thought techniques. This supports the current prompt approaches while suggesting directions to enhance LLMs' causal reasoning capability. Furthermore, CausalBench incorporates background knowledge and training data into prompts to thoroughly unlock LLMs' text-comprehension ability during evaluation, whose findings indicate that, LLM understand causality through semantic associations with distinct entities, rather than directly from contextual information or numerical distributions.

Motivation & Objective

• To address the lack of a comprehensive benchmark for evaluating causal learning in LLMs, which currently suffer from narrow, homogeneous, and simplistic evaluation designs.
• To assess LLMs' performance across a wide range of causal learning tasks, including causal structure identification, correlation vs. causation discrimination, and counterfactual reasoning.
• To investigate how LLMs perform across varying causal network scales (small to large), densities, and structural complexities (e.g., chains, confounders, colliders).
• To examine the impact of background knowledge and structured data on LLMs’ causal reasoning, especially in handling ambiguous variable names and complex inference chains.
• To identify performance gaps between LLMs and state-of-the-art causal learning algorithms, particularly in medium- and large-scale datasets.

Proposed method

• CausalBench constructs a benchmark with three core causal learning tasks: causal structure identification, correlation vs. causation discrimination, and counterfactual reasoning.
• The benchmark includes causal networks of varying sizes (2 to 109 nodes) and densities, enabling evaluation across increasing task difficulty.
• It integrates multiple data modalities: variable names only, variable names with background knowledge, and variable names with structured data (e.g., DAGs or numerical tables).
• The benchmark uses both synthetic and real-world-inspired causal structures to simulate realistic reasoning scenarios, including chains, confounders, and colliders.
• Performance is evaluated using standard metrics: accuracy, F1-score, and causal structure recovery rate, with comparisons to classical and SOTA causal learning algorithms.
• The evaluation includes 19 leading LLMs—both open-source and closed-source—across multiple series of experiments to assess generalization and robustness.

Experimental results

Research questions

• RQ1 How do LLMs perform in distinguishing causal relationships from mere correlations across different network structures and scales?
• RQ2 To what extent does background knowledge improve LLMs’ causal reasoning, especially when variable names are ambiguous?
• RQ3 How does the inclusion of structured data (e.g., DAGs or numerical tables) affect LLMs’ ability to identify causal structures compared to text-only inputs?
• RQ4 How does network complexity—measured by node count, depth, and in-degree/out-degree—impact LLM performance in causal structure identification?
• RQ5 What is the performance gap between LLMs and state-of-the-art causal learning algorithms in medium- and large-scale causal networks?

Key findings

• The highest-performing LLM achieved 65.28% accuracy on small-scale datasets, 74.70% on medium, and 68.06% on large datasets, with F1 scores declining from 0.7017 (small) to 0.5825 (medium) and 0.4310 (large).
• LLMs showed stronger performance on chain structures (F1: ~0.70) but struggled significantly with collider structures (F1: ~0.40), indicating difficulty in handling complex causal dependencies.
• LLM accuracy dropped below 50% when in-degree or out-degree in DAGs exceeded twice the average dataset value, highlighting sensitivity to network sparsity and complexity.
• Background knowledge significantly improved performance only when variable names were semantically clear; it provided little to no benefit for ambiguous names.
• Closed-source LLMs outperformed open-source models in utilizing structured data, while open-source models largely relied on textual cues, limiting their scalability.
• LLMs consistently underperformed compared to SOTA causal learning methods, especially in large-scale datasets, with the performance gap widening as network complexity increased.