Date of Award

Spring 5-28-2026

Document Type

Thesis (Undergraduate)

Department

Computer Science

First Advisor

Peter Chin

Second Advisor

Deeparnab Chakrabarty

Abstract

This thesis utilizes effective rank (erank) to study how attention and MLP sublay-

ers change their residual stream representations in toy transformers. Across the

key-value and modular addition task, we find that the attention/MLP division of

labor is task dependent because the key-value task is heavily attention-sufficient,

while the modular addition transformer heavily relies on both attention and MLP

contributions. During grokking for the modular addition task, late-layer residual

stream representations undergo much more compression than earlier-layer residual

stream representation. This suggests that generalization coincides with late-layer

residual stream compression. Furthermore, we use ablation and freezing experi-

ments to assess component’s significance toward model accuracy and training-time

plasticity. Ablation experiments demonstrate that key-value retrieval has more

localized load-bearing sublayers, while modular addition does not have strong

distinct load-bearing sublayers. Freezing experiments demonstrates that even when

models weights are frozen in layer 0 and cannot compress the residual stream in the

frozen part of the transformer, the model can still grok by moving the compression

to later layers. Together, these results suggest that compressive roles in transformer

residual streams are not fixed to certain sublayers. Instead, compression is task-

dependent and can be redistributed to other sublayers, similar to compensation

mechanism.

Available for download on Saturday, June 05, 2027

Share

COinS