LIBERTy: A Causal Framework for Benchmarking Concept-Based Explanations of LLMs with Structural Counterfactuals

TL;DR

LIBERTy introduces a causal benchmark for evaluating concept-based explanations of LLMs by generating structural counterfactuals from explicitly defined SCMs, thereby overcoming the need for costly human-written counterfactuals. Texts are grounded with exogenous context (persona and templates) and counterfactuals are produced via interventions in the SCM, yielding structured reference targets aligned with the data-generating process. The framework spans three high-stakes datasets (workplace violence, disease detection, and CV screening) and introduces Order-Faithfulness as a robust metric to assess whether explanations preserve the correct ranking of concept effects. Across five explained models and eight explanation methods, LIBERTy finds matching-based explanations perform best locally, while global analyses reveal partial alignment with causal effects and notable room for improvement, with sensitivity analyses showing variation across model families. LIBERTy thus provides a scalable, principled platform for developing and evaluating faithful, causally grounded explainability methods.

Abstract

Concept-based explanations quantify how high-level concepts (e.g., gender or experience) influence model behavior, which is crucial for decision-makers in high-stakes domains. Recent work evaluates the faithfulness of such explanations by comparing them to reference causal effects estimated from counterfactuals. In practice, existing benchmarks rely on costly human-written counterfactuals that serve as an imperfect proxy. To address this, we introduce a framework for constructing datasets containing structural counterfactual pairs: LIBERTy (LLM-based Interventional Benchmark for Explainability with Reference Targets). LIBERTy is grounded in explicitly defined Structured Causal Models (SCMs) of the text generation, interventions on a concept propagate through the SCM until an LLM generates the counterfactual. We introduce three datasets (disease detection, CV screening, and workplace violence prediction) together with a new evaluation metric, order-faithfulness. Using them, we evaluate a wide range of methods across five models and identify substantial headroom for improving concept-based explanations. LIBERTy also enables systematic analysis of model sensitivity to interventions: we find that proprietary LLMs show markedly reduced sensitivity to demographic concepts, likely due to post-training mitigation. Overall, LIBERTy provides a much-needed benchmark for developing faithful explainability methods.

Figures (5)

• Figure 1: Illustration of LIBERTy: The goal is to evaluate an explanation method $M_{f}$ that explains the impact of changing a concept $C$ (by $\stackrel{\rightarrow}{c}$) on model $f$. Left: The causal graph representing the text generation process. Exogenous noise variables are denoted by $\varepsilon$, while the endogenous variables (in this illustration) are the concepts $A,B,C,Y$, the LLM-generated text $x_{\varepsilon}$, and the model prediction $f(x_{\varepsilon})$. The process of generating a structural counterfactual for the change $\stackrel{\rightarrow}{c}$ is highlighted in red: $C$ is assigned a new value and propagated through the causal graph (with $\varepsilon$ fixed) until the LLM generates the counterfactual $x^{\stackrel{\rightarrow}{c}}_{\varepsilon}$. Right: The explanation $M_{f}(x_{\varepsilon}, \stackrel{\rightarrow}{c})$ is compared against the refrence individual causal concept effect (ICaCE), defined as the difference between $f(x^{\stackrel{\rightarrow}{c}}_{\varepsilon})$ and $f(x_\varepsilon)$.
• Figure 2: LIBERTy Causal Graphs: We show only the concepts (endogenous variables) and the relationships between them. Colored concepts indicate the variables that the explained model is trained to predict (the $Y$). At the bottom, a simplified version is provided. The graphs are grounded in prior literature and studies.
• Figure 3: Global Explainability Results: We report the mean Order-Faithfulness score for global explanations. See Table \ref{['tab:complete_global']} in the Appendix for full results.
• Figure 4: Annotation guidelines for validating concept values and rating coherence, fluency, task relevance, and logical consistency. Example of the CV screening dataset.
• Figure 5: Annotation guidelines for rating the plausibility of a text as a genuine counterfactual of the original.

Theorems & Definitions (4)

• Definition 1: Causal Concept Effect (CaCE) and Individual CaCE (ICaCE)
• Definition 2: Empirical CaCE and ICaCE
• Definition 3: ICaCE Error Distance (ED)
• Definition 4: ICaCE Order-Faithfulness (OF)