Google

A step-level search can be visualized as exploring a tree where the root is the input problem, and the tree expands by generating multiple reasoning steps at each point. Each candidate step can be pruned either due to its low generation likelihood from the language model (similar to standard decoding) or by failing an explicit step-level verification check. The surviving unpruned candidates are then expanded to generate further reasoning steps. This continuous process of expansion and pruning allows the system to focus computational resources on the most promising paths, iterating until a complete reasoning chain forming a final answer is generated or a predefined search limit is reached.

Example of Pruning in a Step-Level Search Tree

A system is tasked with generating a step-by-step solution to a math word problem. It explores multiple reasoning paths simultaneously. Below is a log of its activity for two of these paths, Path Alpha and Path Beta. Based on the log, identify which path was likely discarded (pruned) and explain the specific reason for the pruning, referencing the events in the log.

Analyzing a Search Process

A system is solving a problem by exploring multiple reasoning paths one step at a time. Arrange the following events to show the correct logical sequence that occurs within a single iterative step of this process.

A system is solving a multi-step problem by exploring different reasoning paths. It starts with two initial paths, Path X and Path Y. After the first step, it runs a verification check. Path X's first step is found to be logically sound. Path Y's first step is found to contain a factual error. The system continues to expand only the valid path for two more steps to reach a final answer. Based on this process, what is the most likely outcome?

Learn Before

Related