Aligning Tree-Search Policies with Fixed Token Budgets in Test-Time Scaling of LLMs
Researchers have developed a new tree-search decoding algorithm called Budget-Guided MCTS (BG-MCTS) to optimize large language model (LLM) performance within fixed token budgets. This method dynamically adjusts its search strategy, starting with broad exploration and then focusing on refinement as the token limit approaches. BG-MCTS demonstrates superior performance over existing budget-agnostic methods on mathematical and physics reasoning benchmarks using open-weight LLMs. AI
IMPACT This algorithm could improve the efficiency and effectiveness of LLM deployments in resource-constrained environments.