# AutoCompress: Critical Layer Isolation for Efficient Transformer Compression

arXiv: https://arxiv.org/abs/2604.22786 · HTML: https://mlsanity.com/paper/2604.22786/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 7 of its 10 references that are in the map.
- First posted on arXiv 4 Apr 2026.

## Builds on (most cited first, 7 of 7)

- [Distilling the Knowledge in a Neural Network](https://mlsanity.com/paper/1503.02531/) (2015)
- [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://mlsanity.com/paper/1910.01108/) (2019)
- [Pointer Sentinel Mixture Models](https://mlsanity.com/paper/1609.07843/) (2016)
- [TinyBERT: Distilling BERT for Natural Language Understanding](https://mlsanity.com/paper/1909.10351/) (2019)
- [Are Sixteen Heads Really Better than One?](https://mlsanity.com/paper/1905.10650/) (2019)
- [TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate](https://mlsanity.com/paper/2504.19874/) (2025)
- [Iterative Layer-wise Distillation for Efficient Compression of Large Language Models](https://mlsanity.com/paper/2511.05085/) (2025)

## Abstract (arXiv)

We present AutoCompress, a transformer compression method motivated by an empirical finding: in small transformers, Layer 0 carries disproportionately high task-critical information, with an NTK-based importance score of 3.6 compared to a maximum of 0.054 for all other layers -- a gap of over 60x. Based on this finding, we propose Critical Layer Isolation (CLI), an architecture that protects Layer 0 at full dimensionality, compresses all intermediate layers through a learned bottleneck, and restores the full dimension at the final layer. Applied to GPT-2 Medium (354.8M parameters), CLI-GPT2 achieves 204.5 perplexity on WikiText-103 with only 143.8M parameters -- a 2.47x compression ratio and 59.5% parameter reduction. Crucially, an ablation study demonstrates that a uniform bottleneck baseline of comparable size achieves only 571.8 perplexity under identical training conditions, confirming that the architectural decision to protect Layer 0 -- rather than simply reducing model size -- is the primary driver of performance. Code and checkpoints are publicly available.