부산대학교 도서관

How do language models deal with the limited bandwidth of the residual stream? Prior work has suggested that some attention heads and MLP layers may perform a "memory management" role. That is, clearing residual stream directions set by earlier layers by reading in information and writing out the negative version. In this work, we present concrete evidence for this phenomenon in a 4-layer transformer. We identify several heads in layer 2 that consistently remove the output of a single layer 0 head. We then verify that this erasure causally depends on the original written direction. We further demonstrate that direct logit attribution (DLA) suggests that writing and erasing heads directly contribute to predictions, when in fact their effects cancel out. Then we present adversarial prompts for which this effect is particularly salient. These findings reveal that memory management can make DLA results misleading. Accordingly, we make concrete recommendations for circuit analysis to prevent interpretability illusions.