CausalBench: A Comprehensive Benchmark for Evaluating Causal Reasoning Capabilities of Large Language Models

Zeyu WangACL 2024

TL;DR

This paper introduces CausalBench, a novel benchmark designed to rigorously evaluate the causal reasoning abilities of Large Language Models (LLMs) across textual, mathematical, and coding domains. By assessing four distinct causal perspectives (cause-to-effect, effect-to-cause, and their intervention variants), CausalBench aims to move beyond superficial pattern matching and ensure LLMs truly understand causal relationships, a crucial step towards more reliable and intelligent AI.

What Problem Does This Paper Address?

Current evaluations of causal reasoning in LLMs are often limited. Existing benchmarks tend to focus on narrow domains and primarily assess "cause-to-effect" reasoning through text-based problems. This approach doesn't comprehensively test an LLM's understanding, as models might "guess" correct answers without grasping the underlying causal structures. Furthermore, the lack of diverse problem types (like math and code) and a multi-dimensional evaluation strategy means LLMs might appear more capable than they truly are. This paper tackles this by proposing a more robust and comprehensive benchmark to accurately gauge LLM causal reasoning.

Methods

The creation of CausalBench involved a multi-stage process:

  • Manual Generation of Initial Scenarios: The authors began by manually creating test cases covering four causal perspectives: cause-to-effect, effect-to-cause, cause-to-effect with intervention, and effect-to-cause with intervention. This ensured a foundational understanding of the desired problem structures.
  • Scaling Up with LLMs: Leveraging GPT-4 Turbo, the authors expanded the dataset significantly. They used few-shot prompting with the manually created cases to guide GPT-4 Turbo in generating a vast number of additional problems across text, code, and math domains, ensuring consistency with the initial examples.
  • Quality Control: A rigorous quality control process was implemented. This included:
  • Causal Inference Engine: A custom engine was developed to automatically verify the logical consistency and correctness of generated questions and answers by constructing causal graphs and calculating conditional probabilities.
  • Human Expert Review: Human experts reviewed the generated data, especially in cases of discrepancies identified by the engine, to ensure high standards of quality and relevance.

Datasets

  • CausalBench (New Benchmark): This is the primary contribution of the paper. It comprises over 60,000 problems spanning textual, mathematical, and coding domains, each evaluated across four causal reasoning perspectives.
  • CLADDER Dataset: Used as a starting point for generating text-based scenarios.

Key Insights and Ideas

  • The Four-Dimensional Evaluation is Crucial: Requiring LLMs to answer correctly across all four causal perspectives (cause-to-effect, effect-to-cause, and their intervention variants) is key to distinguishing genuine understanding from superficial pattern matching.
  • Domain Differences in Causal Reasoning: LLMs generally perform better on mathematical causal reasoning tasks compared to textual or coding tasks, suggesting that structured, symbolic domains might be more amenable to current LLM capabilities for this type of reasoning.
  • Interventions are Challenging: While interventions are critical for true causal understanding, their introduction sometimes led to decreased performance for LLMs, highlighting a specific area of difficulty.
  • Causal Reasoning and Hallucinations are Linked: A significant finding is the observed correlation between stronger causal reasoning abilities and lower hallucination rates in LLMs. This suggests that a deeper understanding of causality might inherently lead to more factual and less fabricated outputs.

Key Results

  • LLM Performance Varies: State-of-the-art LLMs like GPT-4 and Claude-3 generally outperformed others across the board.
  • Math Outperforms Text/Code: GPT-4 achieved an 88.7% overall correction rate on math problems, compared to 73.3% on text and 71.0% on code.
  • Four-Type Group Correction Rate is Lower: For GPT-4 on math problems, the Four-Type Questions Group Correction Rate was 61.4%, significantly lower than its overall correction rate of 88.7%, indicating struggles with comprehensive scenario understanding.
  • Intervention Impact Varies: In text domains, interventions often decreased correction rates, while in math, they sometimes improved them (e.g., GPT-4's cause-to-effect with intervention on math problems increased from 78.6% to 91.7%).
  • Hallucination Correlation: Models with superior causal reasoning (e.g., GPT-4 Turbo, LLAMA-3-70B) tended to have lower hallucination rates (e.g., GPT-4 Turbo at 2.5%), while those with weaker causal reasoning (e.g., Google Gemma-7b-it) had higher rates (7.5%).

Limitations and Future Work

  • Domain Coverage: The authors acknowledge the need for expanded domain coverage beyond the current text, math, and code.
  • Scale: While significantly larger than previous benchmarks, further increasing the scale of the dataset is desirable.
  • Causal Discovery: The benchmark currently focuses on evaluating reasoning given known causal structures. Future work could incorporate causal discovery tasks.
  • Intrinsic Mechanisms: Deeper exploration into the underlying mechanisms connecting causal reasoning and hallucination reduction is suggested.
  • Empirical Studies: More empirical studies are needed to fully understand the relationship between causal reasoning and hallucination.

Summary updated

BibTeX
@inproceedings{wang2024causalbench,
  author = {Zeyu Wang},
  title = {{CausalBench: A Comprehensive Benchmark for Evaluating Causal Reasoning Capabilities of Large Language Models}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
}
Figure