The Diminishing Returns of Early-Exit Decoding in Modern LLMs

TL;DR

This paper reevaluates early-exit decoding in modern large language models, revealing diminishing benefits in newer models and highlighting factors influencing early-exit potential.

Contribution

It introduces a metric and benchmark for assessing early-exit suitability, and analyzes how model architecture and size affect early-exit effectiveness.

Findings

Early-exit benefits diminish in newer LLMs.

Dense transformers outperform Mixture-of-Experts and State Space Models for early-exit.

Larger models (>20B parameters) show higher early-exit potential.

Abstract

In Large Language Model (LLM) inference, early-exit refers to stopping computation at an intermediate layer once the prediction is sufficiently confident, thereby reducing latency and cost. However, recent LLMs adopt improved pretraining recipes and architectures that reduce layer redundancy, potentially limiting early-exit opportunities. We re-evaluate layer-wise early-exit in modern LLMs and analyze how intermediate representations evolve during training. We introduce a metric to quantify a model's intrinsic suitability for early-exit and propose a benchmark for researchers to explore the potential early-exit benefits on different models and workloads. Our results show a diminishing trend in early-exit effectiveness across newer model generations. We further find that dense transformers generally offer greater early-exit potential than Mixture-of-Experts and State Space Models. In addition, larger models, particularly those with more than 20 billion parameters, and base pretrained models without specialized tuning tend to exhibit higher early-exit potential.