Agentic Search for Counterfactual Recourse under Fixed LLM Budgets
Researchers have developed a new framework called Comp-MCTS to efficiently generate multiple actionable counterfactual explanations for unfavorable decisions made by predictive models. This method addresses the computational cost of using large language models (LLMs) by optimizing the number of LLM calls within a fixed budget. Experiments on real-world datasets demonstrate that Comp-MCTS significantly outperforms existing baselines in producing unique and validated counterfactuals, offering a favorable balance between quantity, quality, and efficiency. AI
IMPACT Improves the explainability of AI decisions by providing actionable recourse options within computational constraints.