Agentic Search for Counterfactual Recourse under Fixed LLM Budgets

A new framework called Comp-MCTS aims to help people find multiple actionable alternatives when a predictive model issues an unfavorable decision, all while staying within a strict budget for calls to large language models, according to research submitted to arXiv in 2026 [1]. Counterfactual recourse is the practice of identifying feature changes a person could make to reverse a negative automated decision [2]. The researchers note that individuals often benefit from having several feasible options rather than a single explanation [2]. Large language models, which are neural networks trained on vast text corpora to perform tasks such as generation and analysis, can propose these alternatives [8]. However, each prompt to an LLM carries a computational and economic cost, making the number of calls the dominant constraint in practice [2]. The study reframes the challenge as a fixed-budget search problem: generating a set of oracle-validated counterfactuals without exceeding a predetermined limit on LLM queries [2]. Comp-MCTS addresses this through an agentic tree-search process that combines LLM-based proposal generation, validation by an oracle, and compression-guided pruning to focus the budget on novel intervention directions [2]. The approach operates in a training-free, oracle-only setting [2]. Experiments on four real-world tabular datasets showed that Comp-MCTS substantially outperformed single-candidate baselines styled after the LATS framework in the yield of unique, oracle-validated counterfactuals [2]. Against stronger multi-candidate variants, Comp-MCTS offered comparable or higher yield at similar or lower oracle-evaluation cost on three of the four datasets, along with competitive proximity, sparsity, and novelty [2]. The paper was posted on arXiv, an open-access repository that hosts e-prints across disciplines including computer science and statistics and has grown to receive about 24,000 submissions per month as of late 2024 [6]. The work appears under the machine learning category and was submitted on June 7, 2026 [1].