Layer by Layer: Uncovering Hidden Representations in Language Models

TL;DR

This paper reveals that intermediate layers in large language models often encode richer, more useful representations than final layers, challenging conventional practices and suggesting new ways to utilize these hidden features for improved performance.

Contribution

The study introduces a unified framework for evaluating layer representations and demonstrates the superior quality of intermediate layer embeddings across diverse models and tasks.

Findings

Intermediate layers often outperform final layers in downstream tasks.

A new framework quantifies representation quality using information theory and geometry.

Mid-depth embeddings can lead to more robust and accurate models.

Abstract

From extracting features to generating text, the outputs of large language models (LLMs) typically rely on the final layers, following the conventional wisdom that earlier layers capture only low-level cues. However, our analysis shows that intermediate layers can encode even richer representations, often improving performance on a range of downstream tasks. To explain and quantify these hidden-layer properties, we propose a unified framework of representation quality metrics based on information theory, geometry, and invariance to input perturbations. Our framework highlights how each layer balances information compression and signal preservation, revealing why mid-depth embeddings can exceed the last layer's performance. Through extensive experiments on 32 text-embedding tasks across various architectures (transformers, state-space models) and domains (language, vision), we demonstrate that intermediate layers consistently provide stronger features, challenging the standard view on final-layer embeddings and opening new directions on using mid-layer representations for more robust and accurate representations.