Tree search is a common way to improve multi-step reasoning in large language models, but most methods give every step the same expansion budget. A new preprint on arXiv argues that this ignores the marginal utility of expanding a particular step, and that budget should be allocated based on uncertainty instead.

The authors say they first demonstrate this issue across existing search methods, though the abstract cuts off before the details of that demonstration. The proposed approach is 'uncertainty-aware budget allocation,' which would direct computational effort toward steps where expansion is likely to matter most.

Because only the abstract was provided, the specific method, experiments, and results are not available in this summary. No other sources were supplied, so there is no independent comparison to report.