← Back to Research

October 30, 2025 · 1 min read

Sparse Autoencoders for Circuit Discovery

Research notes on using SAEs to decompose transformer activations into interpretable features.

interpretability

transformers

research

Sparse autoencoders (SAEs) decompose dense neural network activations into sparse, interpretable features.

Method

class SparseAutoencoder(nn.Module):
    def forward(self, x):
        features = F.relu(self.encoder(x - self.pre_bias) + self.encoder_bias)
        reconstructed = self.decoder(features) + self.pre_bias
        return reconstructed, features

Findings

Feature sparsity of ~95% produces the most interpretable features
Layer 6-8 in GPT-2 small contain the most semantically meaningful features
Feature steering confirms causal relevance for ~60% of top features