Agentic Search for Counterfactual Recourse under Fixed LLM Budgets

Researchers propose Comp-MCTS, an AI framework that efficiently generates multiple counterfactual explanations under limited LLM budget constraints by using tree-search algorithms to allocate queries toward novel intervention directions. The approach demonstrates superior performance in producing diverse, validated counterfactuals compared to existing single-candidate and multi-candidate baselines on real-world datasets.

This research addresses a practical problem at the intersection of explainable AI and economic constraints: generating multiple actionable explanations for algorithmic decisions while minimizing expensive LLM calls. The framework tackles counterfactual recourse, which helps individuals understand what feature changes would reverse unfavorable model decisions. Rather than seeking one optimal explanation, the authors recognize that real-world stakeholders benefit from exploring multiple feasible alternatives—a constraint that becomes computationally expensive when relying on LLM-based generation and validation.

The problem has grown increasingly relevant as organizations deploy LLM-powered AI systems in high-stakes domains like credit decisions, hiring, and lending. These systems demand explainability, yet the cost of generating multiple explanations through LLM prompting creates a practical bottleneck. Comp-MCTS addresses this by framing counterfactual generation as a fixed-budget search problem, employing tree-search methods borrowed from game-playing AI to allocate LLM calls strategically. The approach uses compression-guided pruning to eliminate redundant paths and focuses exploration toward genuinely novel intervention directions.

Experiments across four tabular datasets show Comp-MCTS achieves higher yield of unique, validated counterfactuals while maintaining comparable or lower oracle-evaluation costs than stronger multi-candidate baselines. This efficiency gain matters for deployment: organizations can provide users with more alternatives without proportionally increasing costs. The training-free, oracle-only design also makes the approach flexible across different model types and validation schemes, lowering implementation barriers for enterprises seeking to improve model transparency while controlling computational spending.

• →Comp-MCTS uses tree-search algorithms to efficiently generate multiple counterfactual explanations under fixed LLM-call budgets.
• →The framework outperforms single-candidate and multi-candidate baselines in yield of unique, oracle-validated counterfactuals.
• →Strategic budget allocation via compression-guided pruning reduces redundancy while exploring novel intervention directions.
• →Training-free design allows flexible deployment across different models and validation schemes.
• →Efficiency gains enable organizations to provide users more decision alternatives without proportionally increasing costs.