Preprint reports up to 90% lower token use for LLM tree search

Researchers report a training-free method that reduces token usage by up to 90% and latency by up to 9× at comparable accuracy in tests of large language model tree search. Their preprint introduces Uncertainty-Aware Allocation, which adjusts how many possible reasoning paths a model explores. The method uses differences in meaning among candidate next steps to estimate whether further exploration will produce redundant answers, without additional model cost. It separates choosing promising candidates from deciding how many to expand. The reported results span mathematical reasoning, logical reasoning and planning benchmarks, four tree search algorithms, and both open- and closed-source models.