DiCE vs CARLA: Counterfactual Explanation Libraries

Microsoft's DiCE excels at generating diverse counterfactual explanations because it explicitly optimizes for variety in its loss function. For example, when explaining a denied loan, DiCE can produce multiple distinct paths to approval—'increase income by $5k' and 'reduce debt-to-income ratio by 3%'—rather than repeating a single suggestion. This diversity is critical for public sector applications where citizens need to understand a range of feasible options to contest or change an automated decision.

The CARLA benchmark takes a different approach by evaluating counterfactuals not just on their mathematical properties, but on their actionability and recourse viability. CARLA scores explanations based on whether they suggest changes a citizen can realistically make, penalizing immutable features like age or zip code. This results in a trade-off: CARLA's framework prioritizes real-world usefulness over the theoretical diversity that DiCE maximizes, making it a stricter test for whether an explanation actually empowers a citizen.

The key trade-off: If your priority is generating a wide range of exploratory counterfactuals to help citizens understand decision boundaries, choose DiCE. If you prioritize evaluating and ensuring that the explanations you provide are genuinely actionable and avoid suggesting impossible changes, choose CARLA as your benchmarking standard. For high-stakes government decisions, a combined approach—using DiCE for generation and CARLA for validation—often provides the most robust transparency framework.

DiCE excels at generating a diverse set of counterfactual explanations because it explicitly optimizes for variety using determinantal point processes. For example, a citizen denied a benefit might receive multiple actionable paths—'increase income by $2,000' and 'reduce assets by $5,000'—rather than a single, potentially unachievable suggestion. This diversity is critical for public sector applications where citizens need alternative routes to a positive outcome, directly supporting the 'right to explanation' under regulations like the EU AI Act.

CARLA takes a different approach by providing a standardized benchmarking framework to evaluate the quality of algorithmic recourse. Instead of generating explanations itself, CARLA tests how realistic, actionable, and cost-effective the counterfactuals from any library are using metrics like IM1 (individual recourse cost) and yNN (nearest neighbor realism). This results in a trade-off: CARLA doesn't give you explanations out of the box, but it gives you a rigorous, data-backed method to prove to auditors and legal teams that your explanations are feasible and not just mathematically convenient.

The key trade-off: If your priority is deploying a system that provides citizens with multiple, actionable ways to contest an automated decision immediately, choose DiCE. Its native diversity mechanism directly addresses the need for meaningful algorithmic recourse. If you prioritize validating and auditing the realism of explanations across your AI portfolio to satisfy legal compliance leads, choose CARLA. It serves as the benchmark layer that ensures your chosen explanation method—whether DiCE or another—produces truly actionable results in the real world.