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

Abstract: Causal reasoning is a core component of intelligence. LLMs have shown impressive capabilities in generating human-like text, raising questions about whether their responses reflect true understanding or statistical patterns. We compared causal reasoning in humans and four LLMs using tasks based on collider graphs, rating the likelihood of a query variable occurring given evidence from other variables. LLMs' causal inferences ranged from often nonsensical (GPT-3.5) to human-like to often more normatively aligned than those of humans (GPT-4o, Gemini-Pro, and Claude). Computational model fitting showed that one reason for GPT-4o, Gemini-Pro, and Claude's superior performance is they didn't exhibit the "associative bias" that plagues human causal reasoning. Nevertheless, even these LLMs did not fully capture subtler reasoning patterns associated with collider graphs, such as "explaining away".

• The paper demonstrates that LLMs can perform causal reasoning comparable to or even beyond human normative standards using collider causal structures.
• It employs a comparative methodology with Spearman correlations across predictive, independence, and explaining away inferences between various LLMs and human data.
• The study reveals that LLM performance varies by model and domain knowledge, emphasizing the influence of training data on causal reasoning abilities.

Do LLMs Reason Causally Like Us? Even Better?

Introduction

The research examines the causal reasoning abilities of LLMs, specifically evaluating how they compare to human reasoning regarding causal structures. Causal reasoning involves understanding relationships between variables beyond mere correlations, and it's essential for tasks such as policy recommendations and disease diagnosis. This paper explores the proficiency of LLMs in performing causal reasoning tasks akin to human reasoning.

Methodology

Participants and Models

The study involved comparing human data from Rehder (2017), consisting of 48 undergraduate participants, with outputs from four LLMs: GPT-3.5, GPT-4o, Claude, and Gemini-Pro. The models were prompted with inference tasks over five temperature settings, although only results for a temperature of 0.0 were detailed for consistency.

Materials and Procedure

A collider causal structure ($C_1 \rightarrow E \leftarrow C_2$) was used within three domains: meteorology, economics, and sociology. Each domain presented tasks involving predictive and diagnostic inferences, assessing inferences like predictive causality, independence of causes, and explaining away phenomena. Figures illustrate these tasks (Figure 1), showcasing the understanding required for causal inferences.

Figure 1: Visualization of Causal Mechanism per Domain. The leftmost graph represents task X from the diagnostic inference group.

Results

The analysis revealed distinct reasoning patterns across the four LLMs and humans. Comparative evaluation of the models' performances, measured by Spearman correlations between their inferences and human inferences, showed significant alignment.

Inference Patterns

• Predictive Inference: Similar to humans, LLMs recognized that causes increase the likelihood of effects, showcasing fundamental causal reasoning (Figure 2).

Figure 2: Reference Graph (task II).

• Independence of Causes: While Claude exhibited minimal independence violations, GPT-3.5 and Gemini-Pro showed greater associative biases.
• Explaining Away: GPT-4o demonstrated the strongest ability to explain away effects, while Gemini-Pro and GPT-3.5 did not, indicating an associative rather than causative understanding.

Performance metrics further illustrated that the variability in LLM outputs was more pronounced than in humans, suggesting domain knowledge influences LLM reasoning.

Figure 3: The parameter values from the 4-parameter causal Bayes net fits. Error bars are standard deviations.

Discussion

This study indicates that LLMs achieve a level of causal reasoning relevant to human capabilities but with variations attributable to domain knowledge embedded in their training. Different LLMs displayed unique propensities toward associative and normative reasoning, exemplified by varied response ranges.

Claude and GPT-4o were notable for high normative inference correlations, surpassing the alignment demonstrated by humans. Conversely, Gemini-Pro and GPT-3.5 exhibited less alignment with normative models, highlighting limitations in causal comprehension.

The study supports using causal Bayes nets to fit LLM inferences to a normative standard, revealing the extent to which LLMs utilize domain knowledge. Additionally, the ability of some LLMs (like GPT-4o) to deviate from associative reasoning underpins advancements in developing AI systems with improved causal reasoning abilities.

Conclusion

The research indicates that while LLMs like GPT-4o and Claude align closely with normative causal reasoning standards, others such as Gemini-Pro still exhibit significant associative tendencies. This variation underscores the potential for these models to support decision-making processes, although recognizing domain knowledge's impact remains crucial. Future research should broaden causal structure complexity and deepen analysis on domain knowledge effects.