PAC-MCTS: Bias-Aware Pruning for Robust LLM-Guided Search and Planning
Researchers have introduced PAC-MCTS, a novel bias-aware pruning framework designed to enhance the robustness and efficiency of Large Language Model (LLM)-guided search and planning. As autonomous reasoning tasks deepen, candidate action spaces expand exponentially, often exceeding computational limits. Existing heuristic pruning methods lack formal safety guarantees due to systematic biases in LLM evaluators. This study formulates node expansion as a localized Best-Arm Identification problem under bounded bias, deriving sample complexity bounds that identify conditions for safe elimination. The proposed PAC-MCTS framework dynamically adapts confidence bounds during search to mitigate these biases. Experimental results on Blocksworld and ALFWorld benchmarks demonstrate significant improvements over strong baselines, achieving up to 78% fewer API evaluations and more than three times higher sample efficiency under strict compute budgets. Ablation studies further validate the theoretical predictions regarding performance degradation as evaluator bias increases. This work provides critical insights into managing biased exploration in AI systems, offering a structured approach to balancing computational costs with safety guarantees in complex planning scenarios.
Wire timeline
PAC-MCTS: Bias-Aware Pruning for Robust LLM-Guided Search and Planning
Researchers have introduced PAC-MCTS, a novel bias-aware pruning framework designed to enhance the robustness and efficiency of Large Language Model (LLM)-guided search and planning. As autonomous reasoning tasks deepen, candidate action spaces expand exponentially, often exceeding computational limits. Existing heuristic pruning methods lack formal safety guarantees due to systematic biases in LLM evaluators. This study formulates node expansion as a localized Best-Arm Identification problem under bounded bias, deriving sample complexity bounds that identify conditions for safe elimination. The proposed PAC-MCTS framework dynamically adapts confidence bounds during search to mitigate these biases. Experimental results on Blocksworld and ALFWorld benchmarks demonstrate significant improvements over strong baselines, achieving up to 78% fewer API evaluations and more than three times higher sample efficiency under strict compute budgets. Ablation studies further validate the theoretical predictions regarding performance degradation as evaluator bias increases. This work provides critical insights into managing biased exploration in AI systems, offering a structured approach to balancing computational costs with safety guarantees in complex planning scenarios.
cs.AI updates on arXiv.org