Source-linked AI summary
A Critical Review of Causal Reasoning Benchmarks for Large Language Models
Linying Yang, Vik Shirvaikar, Oscar Clivio, Fabian Falck
TL;DR
Many LLM causal-reasoning benchmarks may be solved by retrieving domain knowledge, and many assess only associations rather than higher-level causal reasoning. This review surveys and classifies benchmarks, finds that retrieval and evaluation-design concerns persist even as some tasks add complexity, and proposes criteria for stronger assessment.
Problem
Many benchmarks may reward retrieval of domain knowledge or test only associations, leaving their ability to assess causal reasoning in question.
Source: Introduction · p. 1Method
The review surveys LLM causal-reasoning datasets and tasks, maps them onto causal hierarchies, and derives benchmark-design criteria.
Source: Introduction · p. 1, Review of existing work, datasets and… · p. 1Results
Many tasks rely on pre-existing knowledge or address only basic associations, while some more complex tasks move beyond simple causal-relation identification.
Source: Introduction · p. 1, Benchmarks for lower levels of causal… · p. 2, Contextual reasoning and graph-based… · p. 3Takeaways & Limitations
The review proposes desirable criteria including causal rather than correlative reasoning, open-endedness, scalability, and non-retrievability.
Source: Conclusion and future work · p. 5Standard NLP similarity metrics for generated explanations do not necessarily establish that an explanation accurately reveals causality.
Source: Many benchmarks suffer from key design… · p. 4
Abstract
from arXiv · showhide
Numerous benchmarks aim to evaluate the capabilities of Large Language Models (LLMs) for causal inference and reasoning. However, many of them can likely be solved through the retrieval of domain knowledge, questioning whether they achieve their purpose. In this review, we present a comprehensive overview of LLM benchmarks for causality. We highlight how recent benchmarks move towards a more thorough definition of causal reasoning by incorporating interventional or counterfactual reasoning. We derive a set of criteria that a useful benchmark or set of benchmarks should aim to satisfy. We hope this work will pave the way towards a general framework for the assessment of causal understanding in LLMs and the design of novel benchmarks.
Introduction
The paper reviews LLM causal-reasoning benchmarks, maps their tasks onto causal hierarchies, and proposes criteria for useful evaluation. It notes that many existing tasks assess only association rather than higher-level causal reasoning.
Introduction
- The review maps existing causal-reasoning tasks onto a hierarchy and proposes criteria for benchmarks that evaluate LLM causal reasoning.Its framing draws on existing taxonomies and hierarchies of causal reasoning.Source: Introduction · p. 1
- Zhang et al. (2023) distinguish identifying causal relationships using domain knowledge, discovering knowledge from data, and quantitatively estimating action consequences.The examples range from predicting what happens to a glass after a quick step to estimating a medication dose from outcomes.Source: Introduction · p. 1
- The ladder of causation separates seeing associations, doing interventions, and imagining counterfactual alternatives.These are respectively described through statistical distributions, do-calculus, and scenarios that may contradict what actually happened.Source: Introduction · p. 1
- Many existing tasks reach only the first rung, testing association identification without interventions or counterfactuals.The review aligns this ladder with Zhang et al.'s hierarchy as a foundation for discussing causal reasoning.Source: Introduction · p. 1
Review of existing work, datasets and tasks
The review surveys causal-reasoning datasets and argues that many assess knowledge retrieval or narrow associations rather than higher-level causal reasoning. It also examines newer contextual and graph-based tasks and identifies design problems that complicate evaluation.
Review of existing work, datasets and tasks
- The authors surveyed review papers, dataset repositories, and websites, compiling 39 datasets and tasks used to assess LLM causal reasoning.Source: Review of existing work, datasets and… · pp. 1–2
Benchmarks for lower levels of causal reasoning
- Many existing tasks benchmark “causal parrots”: multiple-choice cues and restricted options can reward retrieval or language correlations without demonstrating causal reasoning.Source: Benchmarks for lower levels of causal… · p. 2
- Commonsense tasks test everyday knowledge, but success—especially without supporting context—does not establish abstraction or counterfactual reasoning.Source: Benchmarks for lower levels of causal… · p. 2
Contextual reasoning and graph-based tasks move towards more complex conceptions of causality
- Story-based tasks raise complexity with multiple events and fictional details, while FCR adds event extraction, fine-grained causality, and why- and what-if questions.Source: Contextual reasoning and graph-based… · p. 3
- Temporal or spatial order can be mistaken for causality, while graph tasks may simplify reasoning through isolated edges or algebraic structures that lack real-world interpretation.Source: Contextual reasoning and graph-based… · pp. 3–4
Many benchmarks suffer from key design issues
- CEG evaluates generated explanations with lexical-overlap metrics such as ROUGE-L, although semantically equivalent explanations can use different wording and some e-CARE references appear incorrect.Source: Many benchmarks suffer from key design… · p. 4
- Benchmark validity is also challenged by possible training-set memorization, poorly crafted e-CARE choices, and BigBench forecasting items lacking non-causal comparison questions.Source: Many benchmarks suffer from key design… · p. 4
- Some tasks labeled causal instead assess language understanding, reading comprehension, translation, or polarity rather than causal relationships.Source: Many benchmarks suffer from key design… · pp. 4–5
Conclusion and future work
The review proposes four desirable criteria for causal-reasoning benchmarks and presents its work as a step toward a Causal Language Understanding Evaluation framework. These criteria call for causal rather than correlational reasoning, open-ended responses, scalability, and contexts that limit retrieval.
Conclusion and future work
- The review presents its criteria and synthesis as a step toward a CLUE framework comprising a minimal but exhaustive set of tasks for assessing causal reasoning.It draws an analogy to GLUE, a framework for general NLP evaluation.Source: Conclusion and future work · p. 5
- Benchmarks should test interventions and/or counterfactuals in causal language, rather than measuring correlations alone.The authors describe these as desirable criteria; they note that not all four are necessary.Source: Conclusion and future work · p. 5
- Open-ended answers are preferred because multiple-choice options can let models match choices to the premise instead of demonstrating causal reasoning.The review notes that strong performance on existing tasks may reflect data-processing and retrieval capabilities.Source: Conclusion and future work · p. 5
- Benchmarks should scale complexity by adding multiple factors while keeping the causal structure consistent.The review suggests examining how performance changes as factors are added; it cites causal chains and natural word chain tasks as positive examples.Source: Conclusion and future work · p. 5
- Benchmarks should use fictional or non-informative contexts so answers cannot simply be retrieved from likely pretraining data.Real-world stories may permit approximate retrieval of familiar responses rather than causal reasoning.Source: Conclusion and future work · p. 5