Spectral Insights into Data-Oblivious Critical Layers in Large Language Models
This work addresses the need for interpretability and robustness in large language models by providing a method to identify critical layers without data dependency, which is incremental but offers practical benefits for domain adaptation and security.
The paper tackled the problem of identifying intrinsic critical layers in pre-fine-tuned large language models using a data-oblivious approach based on Centered Kernel Alignment, showing that these layers are most affected during fine-tuning and enabling applications like efficient domain adaptation and backdoor defense with up to 40% reduction in attack success rates.
Understanding how feature representations evolve across layers in large language models (LLMs) is key to improving their interpretability and robustness. While recent studies have identified critical layers linked to specific functions or behaviors, these efforts typically rely on data-dependent analyses of fine-tuned models, limiting their use to post-hoc settings. In contrast, we introduce a data-oblivious approach to identify intrinsic critical layers in pre-fine-tuned LLMs by analyzing representation dynamics via Centered Kernel Alignment(CKA). We show that layers with significant shifts in representation space are also those most affected during fine-tuning--a pattern that holds consistently across tasks for a given model. Our spectral analysis further reveals that these shifts are driven by changes in the top principal components, which encode semantic transitions from rationales to conclusions. We further apply these findings to two practical scenarios: efficient domain adaptation, where fine-tuning critical layers leads to greater loss reduction compared to non-critical layers; and backdoor defense, where freezing them reduces attack success rates by up to 40%.