CausalAlign -- Evaluating Causal Alignment Between Humans and LLMs

Overview

CausalAlign is a Python package designed to evaluate the causal reasoning abilities of Large Language Models (LLMs) in comparison to human inference. It provides tools to systematically assess how LLMs reason about causal structures—particularly collider graphs — and whether their judgments align with normative Bayesian inference or human reasoning biases.

This package is based on the work presented in the paper:

"Do Large Language Models Reason Causally Like Us? Even Better?"

Hanna M. Dettki, Brenden M. Lake, Charley M. Wu, Bob Rehder

[arXiv:2502.10215]

Key Features

• Implements the collider causal inference framework to compare LLMs' and human reasoning.
• Provides tools for (Spearman) correlation analysis to measure alignment between human and LLM judgments.
• Fits responses to Causal Bayesian Networks (CBNs) to evaluate normative reasoning.
• Supports API calls for multiple LLMs, including GPT-4o, GPT-3.5, Claude-3-Opus, and Gemini-Pro.
  • for the dataset constructed based on experiment 1 (model-only condition) from Rehder&Waldmann, 2017.
  • dataset: Coming soon

Installation

To install the package, clone the repository and install dependencies:

git clone <https://github.com/hmd101/causalalign.git>
cd causalalign
pip install -e .

If you want to install directly from a public GitHub repository:

pip install git+https://github.com/hmd101/causalalign.git

Usage

_Coming soon_

API calls

_Coming soon_

Required Data Format:

_Coming soon_

Data and Models

CausalAlign provides support for evaluating LLMs across different causal inference tasks. It is particularly apt for datasets building on experiments in Rehder&Waldmann, 2017.

It includes built-in support for:

• Human data from Rehder & Waldmann (2017), experiment 1, model-only condition
• LLM inference responses (GPT-4o, GPT-3.5, Claude-3-Opus, and Gemini-Pro)

Some Benchmarking Results:

Paper Summary: "Do Large Language Models Reason Causally Like Us? Even Better?"

Large Language Models (LLMs) have demonstrated impressive reasoning capabilities, but do they reason causally like humans? This study investigates whether LLMs exhibit human-like causal reasoning, whether they follow normative Bayesian inference, or whether they introduce unique biases.

The study compares human judgments with four LLMs (GPT-4o, GPT-3.5, Claude-3-Opus, and Gemini-Pro) on causal inference tasks based on collider structures (e.g., C1 → E ← C2) embedded in three different knowledge domains (economy, sociology, meteorology) based on the cover stories introduced in Rehder & Waldmann /2017:

Causal Reasoning Varies Across LLMs and Humans

We found that:

• GPT-4o and Claude-3 demonstrated the most normative reasoning, adhering more closely to probability theory than human participants.
• Gemini-Pro and GPT-3.5 leaned towards associative reasoning, relying on statistical correlations rather than strict causal logic.
• All models deviated from expected causal independence, but Claude-3 exhibited the least bias, making its responses the most mathematically consistent.

Interestingly, humans rely on heuristics that sometimes contradict pure probability theory. For example, the "explaining away" effect—where observing one cause makes another less likely—was recognized by AI models but handled inconsistently, depending on their training data.

AI and Humans Approach Causality Differently

A key takeaway is that LLMs do not simply replicate human causal reasoning -- it follows a different logic, possibly due to applying domain knowledge-

• While human judgments remained relatively stable across different domains, AI models adapted their reasoning based on the specific context (e.g., economics vs. sociology).
• GPT-4o treated causal relationships as most deterministic, assuming that certain causes always lead to specific effects, especially in the economy domain for diagnostic reasoning types (E=0).
• Humans, on the other hand, embrace uncertainty, recognizing that possibly other caues, not stated in the prompt may have influence. An alternative explanation could be "lazy reasoning", defaulting to what the experimental setup had set as default (here, 50, where humans were supposed to move a slider on a range from 0-100 when providing their likelihood judgments).

We find that LLMs generally can do the causal inference task and follow relatively similar reasoning patterns as humans do. However, there are domain- and task-speicific differences. This work underscores the need to evaluate LLMs’ reasoning beyond pattern recognition, ensuring they make reliable causal inferences in real-world applications.

For more details, refer to the full paper: [arXiv:2502.10215].

Table: Spearman correlations (rₛ) between human and LLM inferences.

Model	Economy (rₛ)	Sociology (rₛ)	Weather (rₛ)	Pooled across all domains (rₛ)
Claude	0.557	**0.637	0.698	0.631
GPT-4o	0.662	0.572	0.645	0.626
GPT-3.5	0.419	0.450	0.518	0.462
Gemini	0.393	0.297	0.427	0.372

Citation

Please cite paper as:

@misc{dettki2025largelanguagemodelsreason,
  title         = {Do Large Language Models Reason Causally Like Us? Even Better?},
  author        = {Hanna M. Dettki and Brenden M. Lake and Charley M. Wu and Bob Rehder},
  year          = {2025},
  eprint        = {2502.10215},
  archiveprefix = {arXiv},
  primaryclass  = {cs.AI},
  url           = {https://arxiv.org/abs/2502.10215}
}