Ghosted Layers: Unconstrained Activation Alignment for Recovering Layer-Pruned LLMs
Fuente:
arXiv
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866909045994029056 |
|---|---|
| author | Yun, Vincent-Daniel Jo, Junhyuk Karimireddy, Sai Praneeth Lee, Sunwoo |
| author_facet | Yun, Vincent-Daniel Jo, Junhyuk Karimireddy, Sai Praneeth Lee, Sunwoo |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_15491 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Ghosted Layers: Unconstrained Activation Alignment for Recovering Layer-Pruned LLMs Yun, Vincent-Daniel Jo, Junhyuk Karimireddy, Sai Praneeth Lee, Sunwoo Machine Learning Artificial Intelligence Performance 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. |
| title | Ghosted Layers: Unconstrained Activation Alignment for Recovering Layer-Pruned LLMs |
| topic | Machine Learning Artificial Intelligence Performance |
| url | https://arxiv.org/abs/2605.15491 |