CS329A Self-Improving AI Agents, Part 5: Planning and Multi-Step Reasoning

Name: CS329A Self-Improving AI Agents, Part 5: Planning and Multi-Step Reasoning
Uploaded: 2026-09-28T00:00:00.000Z
Duration: 4495 s
Channel: Stanford Online
Description: LATS brings Monte Carlo Tree Search to language agents with LLM-judge scoring; SPRINT fine-tunes reasoners to emit independent plans that run in parallel; SWiRL trains on synthetic multi-step tool-use trajectories and generalizes without executing tools during training.

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

SOStanford Online@stanfordonline

Full transcript

English

Summary：LATS brings Monte Carlo Tree Search to language agents with LLM-judge scoring; SPRINT fine-tunes reasoners to emit independent plans that run in parallel; SWiRL trains on synthetic multi-step tool-use trajectories and generalizes without executing tools during training.

Watch on YouTube

Core points (3)

1LATS combines reasoning, acting, and MCTS with LLM-judge and self-consistency scoring.
2SPRINT cuts sequential tokens by generating independent plans for parallel execution.
3SWiRL trains on synthetic tool trajectories without calling tools during training.