Search papers, labs, and topics across Lattice.
This study investigates the causal necessity of single-token sparse autoencoder (SAE) features across different SAE families by analyzing 3.9 million features from six models. The findings reveal that single-token features are more tightly clustered in decoder space and concentrated in early layers, with significant logit reductions observed upon their ablation in most conditions. Importantly, the research uncovers that cross-family differences in causal roles are more pronounced than within-family effects, suggesting that interpretability claims are influenced by training methodology rather than merely activation functions or model scale.
Single-token SAE features can significantly impact model outputs, with their causal roles varying dramatically across different SAE families.
Sparse autoencoder (SAE) features are used to interpret and steer large language models, yet whether a feature's causal role is stable across SAE families remains untested. Single-token features that activate on one vocabulary item provide the diagnostic case where ground truth permits direct comparison. We analyze 3.9M features across six models and three SAE families using zero-ablation at full layer depth. Single-token features cluster 4.7x tighter in decoder space and concentrate in early layers (Layer 0 in GPT2-Small; L0-L4 in Gemma). Ablating them yields Benjamini-Hochberg-significant logit reductions in 178 of 208 full-layer conditions, with depth controlling whether damage cascades downstream or shapes the output directly. Cross-family causal differences exceed within-family scale effects: on the same base model, GemmaScope and BatchTopK features remain causally anchored, while LlamaScope features are locally redundant. The target token's rank recovers to within 2x baseline 96-98% of the time after the same ablation, and a controlled activation-function comparison reverses sign within the same model, leaving training recipe as the residual candidate. Cross-family interpretability claims are therefore sensitive to training methodology, not just activation function or scale.