Inside Language Models
How can we connect a readable account of a model’s state to the work it actually does?
Start with the first report → See all reports ↓Overview
A model does numerical work before selecting a word. An intermediate state may carry an identity, a relation or an unfinished result into later computation. Making that information readable creates a second problem: which parts did a description preserve, and how do they relate to the model’s behavior?
The recommended order begins with how layers, attention and specially trained latent systems let internal information participate in further work. Then the activation-readout article explains how words can be trained to reconstruct a numerical state, and how edits and interventions test what the words reveal. This order gives the mechanism a foundation before introducing a specialized way of investigating it.
Using an internal state and explaining it are separate achievements
Carry information into later work
Layers and attention let intermediate content affect another computation.
Read it, then test the account
Language makes a state inspectable; controlled changes probe what it explains.
Start with computation, then investigate its readout. Readable content alone does not establish a behavioral explanation.
Reading order
A recommended path through the enquiry. Each report explains its part; continuation links preserve the original branches.
- Understand the work before the wordsHow Language Models Use Ideas Before Putting Them Into Words
Builds the foundation: how numerical representations remain accessible and become inputs to layers, attention and additional latent steps.
Start here - Make an internal state inspectableHow a Model’s Numbers Become Words
Introduces activation-to-language interfaces, then separates readable descriptions, reconstruction and interventions on behavior.