The Inside View cover image

The Inside View

Collin Burns On Discovering Latent Knowledge In Language Models Without Supervision

Jan 17, 2023
02:34:39

Podcast summary created with Snipd AI

Quick takeaways

  • An unsupervised method to identify true or false statements in language models using hidden states as features and training a linear model for classification.
  • Exploring the structure of truth and leveraging logical consistency to discover truth-like features in language models.

Deep dives

Unsupervised discovery of truth in language models

Our paper introduces an unsupervised method to identify whether statements in language models are true or false. We achieve this by using the hidden states of the language model as features and training a linear model to classify statements and their negations as either true or false. The method leverages the structure of truth, including logical consistency and negation consistency, to find features that are correlated with the truth. This approach allows us to identify truth in language models without relying on any explicit supervision or labels, making it scalable and applicable to situations where evaluating truthfulness is challenging.

Remember Everything You Learn from Podcasts

Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.
App store bannerPlay store banner