Ghosted Layers: Unconstrained Activation Alignment for Recovering Layer-Pruned LLMs

TL;DR

Ghosted Layers is a training-free method that recovers pruned large language models by aligning activations, improving performance without retraining.

Contribution

It introduces a closed-form optimal linear operator for activation alignment, surpassing constrained solutions and enhancing pruned LLM accuracy.

Findings

Consistently improves accuracy and perplexity across multiple LLMs.

Outperforms prior training-free baselines.

Preserves efficiency gains of layer pruning.

Abstract

Layer pruning removes entire Transformer decoder blocks from large language models, but introduces a mismatch between the hidden state received by the next surviving layer and the distribution it was trained to process, leading to significant performance degradation. We propose Ghosted Layers, a training-free recovery module that addresses this issue by solving a boundary activation alignment problem. Our method derives a closed-form optimal linear operator from a small calibration set to reconstruct the activation discrepancy introduced by the pruned layers. We show that this solution corresponds to the unconstrained optimum of the alignment objective, whereas existing methods are restricted to constrained solutions over limited operator subspaces. Experiments across multiple LLM backbones and pruning strategies demonstrate that our method consistently improves accuracy and perplexity over prior training-free baselines, while preserving the efficiency gains of layer pruning.