Skip to content
Trending storyDeveloping

MIT course lecture covers RL and self-distillation for LLM training

1 article1 sourcesince Nov 9Last article 20h ago ·

Overview

AISummary of 1 article

MIT's MAS.S62 lecture 6, taught by Idan Shenfeld of Thinking Machines Lab, covers reinforcement learning for large language models using PPO, GRPO and on-policy distillation, according to a post by X user meng shao.

The lecture contrasts on-policy distillation, which assigns a score to each token, with RL, which gives one reward per response. It also describes the five steps of a single training update, lists the training curves to monitor, and shows how to run the verl framework, comparing it with the reef online-learning server.

Written by AI from the articles below · updated Nov 9, 9:27 PM ET

Check the sources:

Latest developmentNov 9, 1:17 AM ET
MIT course lecture walks through RL and self-distillation for LLM training

Article timeline

The articles in this story. Times are ET.

Nov 9
  1. meng shaoX
    MIT course lecture walks through RL and self-distillation for LLM training

    AIMIT's MAS.S62 lecture 6, taught by Idan Shenfeld of Thinking Machines Lab, covers running reinforcement learning on LLMs with PPO, GRPO, and on-policy distillation. The lecture contrasts on-policy distillation, which gives a score per token, with RL's single reward per response, and describes the five steps of one training update. It also lists the training curves to monitor and shows how to run the verl framework, then compares it with the reef online-learning server.

Heat Trend

Not enough continuous observations to show a trend yet.