From Causal Parrots to Causal Prophets? Towards Sound Causal Reasoning with Large Language Models

Rahul Babu Shrestha, Simon Malberg, Georg GrohNAACL 2025

TL;DR

This paper investigates the causal reasoning capabilities of Large Language Models (LLMs) across different levels of causal complexity. They introduce novel prompting techniques, CAUSALTOT and CAUSALPOT, which significantly improve LLM performance, particularly in challenging counterfactual reasoning tasks, suggesting a path towards more robust causal AI.

What Problem Does This Paper Address?

Causal reasoning, the ability to understand cause-and-effect relationships, is a cornerstone of intelligence. While LLMs have shown remarkable progress in language tasks, their ability to perform genuine causal inference beyond mere pattern matching remains a subject of debate. Existing research suggests LLMs might be "causal parrots," regurgitating learned associations rather than truly reasoning. This is particularly problematic in high-stakes domains like healthcare and policy where sound causal understanding is crucial for reliable decision-making. This paper aims to systematically evaluate LLMs' causal reasoning abilities and explore methods to enhance them, moving beyond simple correlations to interventions and counterfactuals.

Methods

The authors evaluate LLMs on causal reasoning tasks using the CLADDER dataset, which covers three levels of Pearl's Ladder of Causation: associational, interventional, and counterfactual. To test robustness, they also introduce anti-commonsensical and nonsensical perturbations to the dataset.

A key contribution is the introduction of two novel prompting techniques:

  • Causal Tree of Thoughts (CAUSALTOT): An adaptation of the "tree of thoughts" approach, CAUSALTOT iteratively generates, evaluates, and selects multiple causal thought candidates at each step of the causal reasoning process. This allows for exploring alternative reasoning paths and choosing the most promising one.
  • Causal Program of Thoughts (CAUSALPOT): This technique integrates LLMs with the DoWhy causal inference library. The LLM generates DoWhy code to perform causal estimations, and the resulting estimate is then fed back to the LLM to derive the final answer. This bridges the gap between LLM reasoning and formal causal inference tools.

They also evaluate standard prompting techniques like Chain of Thought (CoT) and Self-Consistency (SC).

Datasets

The primary dataset used is CLADDER (Causal Reasoning Benchmark). The authors sampled 1,000 commonsensical questions from CLADDER and created new anti-commonsensical and nonsensical versions of these questions using GPT-40 to test model robustness against unfamiliar causal structures.

Key Insights and Ideas

  • LLMs are improving, but counterfactuals remain tough: Newer, larger LLMs generally outperform older ones and are more robust to perturbations. However, counterfactual reasoning consistently proves to be the most challenging rung of the Ladder of Causation for all tested models.
  • Beyond "causal parrots": While some LLMs might rely on learned knowledge, the ability to perform reasonably well on anti-commonsensical and even nonsensical questions suggests an emerging capacity for genuine causal reasoning, especially in larger models.
  • Prompting is key, but specialized prompts unlock more: Standard prompting techniques like CoT offer some improvement, but the newly proposed CAUSALTOT and CAUSALPOT significantly boost performance, particularly CAUSALPOT, which shows consistent accuracy across all causal rungs, including counterfactuals.
  • Formal tools are powerful allies: Integrating LLMs with formal causal inference libraries like DoWhy (as in CAUSALPOT) shows great promise in overcoming LLM limitations in complex causal reasoning.

Key Results

  • Baseline performance: The strongest model, DeepSeek R1, achieved an average accuracy of 89.1% on commonsensical questions using simple input-output prompting.
  • Prompting technique impact:
    • CAUSALTOT improved GPT-40's accuracy by an average of 4.4% points over input-output prompting.
  • CAUSALPOT achieved an even greater improvement, averaging 8.8% points over input-output prompting.
  • CAUSALPOT demonstrated particularly strong performance on counterfactual questions, matching its performance on associational and interventional tasks, a feat not achieved by other methods.
  • Robustness: While smaller models showed a significant performance drop on anti-commonsensical and nonsensical questions, larger models maintained performance, sometimes even showing slight increases, suggesting emergent causal reasoning abilities.

Limitations and Future Work

The authors acknowledge that their study is limited to the CLADDER dataset, which is a synthetic benchmark. They suggest future work should evaluate their proposed methods on other formal causal reasoning datasets to assess generalizability. They also emphasize the need to explore methods that extend beyond formal reasoning to informal reasoning tasks. Furthermore, they highlight the ethical implications of LLM-based causal reasoning, warning against over-reliance and the potential for misuse in high-stakes decision-making.

Summary updated

BibTeX
@inproceedings{shrestha2025from,
  author = {Rahul Babu Shrestha and Simon Malberg and Georg Groh},
  title = {{From Causal Parrots to Causal Prophets? Towards Sound Causal Reasoning with Large Language Models}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure