We introduce methods for discovering and applying sparse feature circuits. These are causally
implicated subnetworks of human-interpretable features for explaining language model …

… In these terms we can examine dependencies of nodes with the output of earlier nodes, ie
we can measure the effect of attention heads in layer 0 on the attention heads in layer 2. In the …

What latent features are encoded in language model (LM) representations? Recent work on
training sparse autoencoders (SAEs) to disentangle interpretable features in LM …

… How can we understand what these computations represent, such that we can arrive at a
deeper algorithmic understanding of how and why models behave the way they do? For …

The enormous scale of state-of-the-art foundation models has limited their accessibility to
scientists, because customized experiments at large model sizes require costly hardware and …

… rage has become a serious concern in many countries, and preventive efforts are required.
This article reviews what can be done to prevent road rage … for road rage behavior could be …

Transformer models underpin many recent advances in practical machine learning applications,
yet understanding their internal behavior continues to elude researchers. Given the size …

… Therefore, if real, memory management could significantly impact what conclusions we can
… 2, we can see the projection ratio between the outputs of every attention head and MLP (…

… can be deployed on mazes smaller than the training size without destroying the structure of
the vocabulary. Examples of usage of this dataset to train autoregressive transformers can be …

… and can be computed in seconds, enabling frequent evaluations such as during SAE training.
However, we can … This dependency on human-generated concepts can overlook important …