Class Central is learner-supported. When you buy through links on our site, we may earn an affiliate commission.

YouTube

Language Model Alignment: Theory and Algorithms

Simons Institute via YouTube

Overview

Explore the intricacies of language model alignment in this comprehensive lecture by Ahmad Beirami from Google. Delve into the post-training process aimed at generating samples from an aligned distribution to enhance rewards such as safety and factuality while minimizing divergence from the base model. Examine the best-of-N baseline technique and more advanced methods solving KL-regularized reinforcement learning problems. Gain insights into key results through simplified examples and discover a novel modular alignment approach called controlled decoding. Learn how this technique solves the KL-regularized RL problem while maintaining a frozen base model through prefix scorer learning, offering inference-time configurability. Analyze the surprising effectiveness of best-of-N in achieving competitive or superior reward-KL tradeoffs compared to state-of-the-art alignment baselines.

Syllabus

Language Model Alignment: Theory & Algorithms

Taught by

Simons Institute

Reviews

Start your review of Language Model Alignment: Theory and Algorithms

Never Stop Learning.

Get personalized course recommendations, track subjects and courses with reminders, and more.

Someone learning on their laptop while sitting on the floor.