The paper "LiteSearch: Efficacious Tree Search for LLM" introduces a novel guided tree search algorithm designed to enhance the performance of large language models (LLMs) on complex mathematical reasoning tasks while significantly reducing computational costs. The authors address the issue of inefficient search strategies in tree search algorithms, such as Monte Carlo Tree Search (MCTS), which often require more computational resources than greedy decoding methods. Their proposed algorithm, LiteSearch, employs dynamic node selection and node-level exploration budget calculation, guided by a value network trained without step-wise annotations. This approach balances exploration and exploitation by iteratively selecting the most promising tree node and expanding it within a dynamically computed budget, leading to faster convergence and lower computational costs. Experiments on the GSM8K and TabMWP datasets demonstrate that LiteSearch offers competitive performance with significantly reduced computation costs compared to baseline methods. The paper also includes a detailed analysis of the algorithm's components and their effectiveness, as well as discussions on limitations and potential future directions.The paper "LiteSearch: Efficacious Tree Search for LLM" introduces a novel guided tree search algorithm designed to enhance the performance of large language models (LLMs) on complex mathematical reasoning tasks while significantly reducing computational costs. The authors address the issue of inefficient search strategies in tree search algorithms, such as Monte Carlo Tree Search (MCTS), which often require more computational resources than greedy decoding methods. Their proposed algorithm, LiteSearch, employs dynamic node selection and node-level exploration budget calculation, guided by a value network trained without step-wise annotations. This approach balances exploration and exploitation by iteratively selecting the most promising tree node and expanding it within a dynamically computed budget, leading to faster convergence and lower computational costs. Experiments on the GSM8K and TabMWP datasets demonstrate that LiteSearch offers competitive performance with significantly reduced computation costs compared to baseline methods. The paper also includes a detailed analysis of the algorithm's components and their effectiveness, as well as discussions on limitations and potential future directions.