Copy Suppression: Comprehensively Understanding an Attention Head

TL;DR

This work identifies copy suppression as the principal function of a single attention head (L10H7) in GPT-2 Small, outlining a three-step mechanism: prior naive copying, attention to the copied source, and direct suppression of the copied token’s logit. Using OV and QK circuit analyses, the authors provide weights-based evidence that most of L10H7’s behavior across the training distribution can be attributed to suppressing tokens it attends to or predicts. They validate this with a KL-divergence–based ablation framework (CSPA), finding that copy suppression accounts for 76.9% of L10H7’s effect, with higher explainability when isolating OV or QK pathways. The study also connects copy suppression to self-repair phenomena, demonstrating that roughly 39% of IOI-task self-repair can be explained by copy-suppressing mechanisms, while acknowledging additional contributing factors beyond unembedding directions. Overall, the work advances mechanistic interpretability by providing a comprehensive, distributed-scope account of a single component’s role and its practical implications for calibration and robust behavior.

Abstract

We present a single attention head in GPT-2 Small that has one main role across the entire training distribution. If components in earlier layers predict a certain token, and this token appears earlier in the context, the head suppresses it: we call this copy suppression. Attention Head 10.7 (L10H7) suppresses naive copying behavior which improves overall model calibration. This explains why multiple prior works studying certain narrow tasks found negative heads that systematically favored the wrong answer. We uncover the mechanism that the Negative Heads use for copy suppression with weights-based evidence and are able to explain 76.9% of the impact of L10H7 in GPT-2 Small. To the best of our knowledge, this is the most comprehensive description of the complete role of a component in a language model to date. One major effect of copy suppression is its role in self-repair. Self-repair refers to how ablating crucial model components results in downstream neural network parts compensating for this ablation. Copy suppression leads to self-repair: if an initial overconfident copier is ablated, then there is nothing to suppress. We show that self-repair is implemented by several mechanisms, one of which is copy suppression, which explains 39% of the behavior in a narrow task. Interactive visualisations of the copy suppression phenomena may be seen at our web app https://copy-suppression.streamlit.app/

Figures (26)

• Figure 1: L10H7's copy suppression mechanism (Steps 1-3 are described below and the mechanism is examined in \ref{['sec:how_negative_heads_copy_suppress']}).$^*$Work partially done at Google DeepMind
• Figure 2: Loss effect of L10H7 via different paths. Grey paths denote ablated paths.
• Figure 3: Distribution of ranks of diagonal elements of .
• Figure 4: Illustration of three different kinds of ablation: just OV, just QK, and CSPA.
• Figure 5: We plot $(\overline{D_{\text{MA}}}, \overline{D_{\text{CSPA}}})$ for each percentile of our OpenWebText data (with percentiles given by the values of $D_\text{MA}$).