Uncertainty-Aware Budget Allocation for Efficient LLM Tree Search

Key points

• We first show, across existing search algorithms, that continuations sampled from a shared intermediate state are largely duplicates and collapse to a few distinct answers.1
• Critically, we find that the semantic dispersion of sibling actions predicts the redundancy of their downstream continuations, permitting a subtree's budget to be fixed before expansion.1
• An information-theoretic analysis motivates this: treating the per-step budget as the Monte Carlo sample size for estimating a step's residual answer entropy, we derive a sufficient budget that shrinks with sibling dispersion.1
• We accordingly propose $Uncertainty-Aware Allocation$ (UAA), a training-free framework that decouples value from uncertainty: values determine which candidates are promising, while uncertainty, estimated from sibling dispersion at no additional model cost, determines how many merit expansion, via uncertainty-aware pruning and uncertainty-aware budgeting.1

Sources (1)

• [1]Uncertainty-Aware Budget Allocation for Efficient LLM Tree Search
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 16, 03:03 AM

  Tree search enhances multi-step reasoning in large language models (LLMs), yet prevailing methods expand every step with a fixed budget, irrespective of the marginal utility of expansion.

  We first show, across existing search algorithms, that continuations sampled from a shared intermediate state are largely duplicates and collapse to a few distinct answers.

Extractive summary: sentences quoted from the sources.