Layer by Layer: Uncovering Hidden Representations in Language Models

Oscar Skean, Md Rifat Arefin, Dan Zhao, Niket Patel, Jalal Naghiyev, Yann LeCun, Ravid Shwartz-Ziv

International Conference on Machine Learning 2025 · Oral

In the rapidly evolving landscape of large language models (LLMs), a set of deeply ingrained assumptions often guides their application: that the final layers yield the most optimal embeddings for downstream tasks, or that intermediate layers are largely unhelpful. This talk, "Layer by Layer: Uncovering Hidden Representations in Language Models," delivered by Oscar Skean, a PhD student at the University of Kentucky, boldly challenges this conventional wisdom. Skean, presenting on behalf of a diverse, cross-institutional team including prominent researchers like Yann LeCun, unveils compelling empirical evidence demonstrating that intermediate layers frequently surpass the performance of final layers across a wide array of models, scales, and modalities.

AI review

A well-executed empirical study demonstrating that intermediate layers of autoregressive transformers consistently outperform final layers on downstream tasks, with prompt entropy (effective rank of the token covariance matrix) proposed as a predictive heuristic. The core observation is real and practically useful, and the cross-architecture and cross-modality comparisons strengthen the case that this is a property of autoregressive training rather than a dataset artifact. However, the theoretical contribution is thinner than the framing suggests: prompt entropy is a repackaging of effective…