Concept

Backtracking in Step-Level Search

Step-by-step verification enables dynamic adjustments to the reasoning process through backtracking. If a candidate step, denoted as ai+1a_{i+1}, is deemed incorrect or unpromising by the verifier V()V(\cdot), the search algorithm can backtrack to explore alternative steps from the immediately preceding partial path ai\mathbf{a}_{\le i}, or even from an earlier node ai\mathbf{a}_{\le i'} (where i<ii' < i). Conversely, computational resources are focused on extending paths that receive high ratings.

0

1

Updated 2026-05-06

Contributors are:

Who are from:

Tags

Foundations of Large Language Models

Ch.5 Inference - Foundations of Large Language Models

Foundations of Large Language Models Course

Computing Sciences