# Wiring Beats Blending: Structure-Aware Compensation for Transformer Downscaling

arXiv: https://arxiv.org/abs/2608.02829 · HTML: https://mlsanity.com/paper/2608.02829/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 15 of its 17 references that are in the map.
- First posted on arXiv 3 Aug 2026.

## Builds on (most cited first, 15 of 15)

- [Distilling the Knowledge in a Neural Network](https://mlsanity.com/paper/1503.02531/) (2015)
- [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://mlsanity.com/paper/2104.09864/) (2021)
- [Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling](https://mlsanity.com/paper/2304.01373/) (2023)
- [Similarity of Neural Network Representations Revisited](https://mlsanity.com/paper/1905.00414/) (2019)
- [LLM-Pruner: On the Structural Pruning of Large Language Models](https://mlsanity.com/paper/2305.11627/) (2023)
- [Git Re-Basin: Merging Models modulo Permutation Symmetries](https://mlsanity.com/paper/2209.04836/) (2022)
- [Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning](https://mlsanity.com/paper/2310.06694/) (2023)
- [Net2Net: Accelerating Learning via Knowledge Transfer](https://mlsanity.com/paper/1511.05641/) (2015)
- [Understanding image representations by measuring their equivariance and equivalence](https://mlsanity.com/paper/1411.5908/) (2014)
- [Revisiting Model Stitching to Compare Neural Representations](https://mlsanity.com/paper/2106.07682/) (2021)
- [Compact Language Models via Pruning and Knowledge Distillation](https://mlsanity.com/paper/2407.14679/) (2024)
- [Learning to Grow Pretrained Models for Efficient Transformer Training](https://mlsanity.com/paper/2303.00980/) (2023)
- [bert2BERT: Towards Reusable Pretrained Language Models](https://mlsanity.com/paper/2110.07143/) (2021)
- [Staged Training for Transformer Language Models](https://mlsanity.com/paper/2203.06211/) (2022)
- [Weight subcloning: direct initialization of transformers using larger pretrained ones](https://mlsanity.com/paper/2312.09299/) (2023)

## Abstract (arXiv)

Model families are trained size by size. Can a pretrained large model instead be converted into a smaller sibling? We study the 1.4B->410M conversion in Pythia end to end. Representations align strongly across sizes (ridge R^2=0.84); parameters align weakly. Dense weight projection is destructive; a bit-exact control places the fault in basis mixing, which breaks rotary, per-head, GELU, and LayerNorm structure. Residuals after the best-fit linear operator carry no learnable or transferable signal under shuffle controls, so conversion value lives in initialization. Matched-budget continued pre-training separates two independent levers: least-squares compensation (function lever, best zero-shot) and variance-preserving rescale (dynamics lever, best endpoints). Placement follows the architecture: compensation is well-posed exactly where no normalization sits between cut and read; norm-fronted paths take rescale. Compensation is a low-budget, token-efficiency win, not a universal one. At 30M tokens it beats the best subcloning variant on a width-reduced pair (84.0+-1.8 vs. 89.7+-3.7, 3/3 seeds) and a held-out depth-reduced pair (109.3 vs. 117.9, 3/3 seeds). Selection given the same activation statistics recovers under half of that gap (3/3 seeds): the gain is the re-fit, not the information. At 33x the budget the two reach parity (40.3+-0.3 vs. 40.3+-0.5, 3 seeds), both far ahead of from-scratch, which transfer always beats (up to 18x at low budget, narrowing at convergence and at the largest scale). At ~5x the donor scale (6.9B->1.4B) stacking both levers over-corrects, consistent with an ill-conditioned compensation solve at large width, pointing to dimension-aware regularization as a fix. The init also beats structured pruning plus distillation, the standard pipeline, at matched budget, and improves further combined with it. Code, checkpoints, and the frozen eval corpus are released.