Vibeleaderboard
← All Intel
Intel / video

Stanford CS329A Self-Improving AI Agents | Part 5 | Planning and Multi-Step Reasoning

Source
youtube.com
Author
Stanford Online
Date

Context

Language Agent Tree Search, or LATS, lets an AI agent try several actions and learn from the results instead of committing to one plan. A Stanford CS329A lecture walks through the paper, which brings Monte Carlo tree search, a method for exploring branching choices by repeatedly sampling them, into language-model agents.

In the lecture's maze example, a model proposes several next actions, each is carried out in the environment, and the result is appended to the . Each new is scored by adding two measures: a 's rating of how promising it is, and a self-consistency score based on how often that action appears across many samples. A rule borrowed from tree search then balances exploring less-visited branches against following the best one. When a ends in success or failure, the outcome updates the scores of earlier steps, and the model writes a reflection.

The lecturer contrasts this with Math-Shepherd, where a trained verifier scores reasoning steps, saying LATS scores action outcomes. The excerpt reports no benchmark results.

Terms in this piece · Glossary
  • Context and context window — Context is the information available for a generation; the context window is the capacity available to hold it.
  • State — Information describing a system at a particular moment, such as its current text, settings or saved records.
  • Large language model — A model trained on large amounts of data to process and generate language.
  • Path — A description of a file or folder location, written using directory names and separators.
Read the source www.youtube.com
More from Stanford Online
Recommended reads
Comments

Checking sign-in…

Loading comments…