Picture for Can Rager

Can Rager

Colorado School of Mines, Department of Applied Mathematics and Statistics

Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models

Add code
Mar 31, 2024
Figure 1 for Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models
Figure 2 for Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models
Figure 3 for Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models
Figure 4 for Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models
Viaarxiv icon

Structured World Representations in Maze-Solving Transformers

Add code
Dec 05, 2023
Viaarxiv icon

Attribution Patching Outperforms Automated Circuit Discovery

Oct 16, 2023
Viaarxiv icon

An Adversarial Example for Direct Logit Attribution: Memory Management in gelu-4l

Add code
Oct 14, 2023
Figure 1 for An Adversarial Example for Direct Logit Attribution: Memory Management in gelu-4l
Figure 2 for An Adversarial Example for Direct Logit Attribution: Memory Management in gelu-4l
Figure 3 for An Adversarial Example for Direct Logit Attribution: Memory Management in gelu-4l
Figure 4 for An Adversarial Example for Direct Logit Attribution: Memory Management in gelu-4l
Viaarxiv icon

A Configurable Library for Generating and Manipulating Maze Datasets

Add code
Sep 19, 2023
Viaarxiv icon