Do Depth-Grown Models Overcome the Curse of Depth? An In-Depth Analysis

TL;DR

This paper investigates how gradually increasing the depth of Transformer models during training enhances their reasoning capabilities by improving depth utilization, restructuring residual streams, and forming modular computational blocks, thus overcoming the traditional curse of depth.

Contribution

It provides a mechanistic understanding of depth growth benefits, linking it to better depth utilization and circuit formation, and proposes a lightweight modification to further improve reasoning performance.

Findings

Depth-wise analysis shows improved utilization of later layers in grown models.

Gradual depth growth alters residual stream structure and creates permutable blocks.

Modified MIDAS achieves better downstream reasoning benchmarks.

Abstract

Gradually growing the depth of Transformers during training can not only reduce training cost but also lead to improved reasoning performance, as shown by MIDAS (Saunshi et al., 2024). Thus far, however, a mechanistic understanding of these gains has been missing. In this work, we establish a connection to recent work showing that layers in the second half of non-grown, pre-layernorm Transformers contribute much less to the final output distribution than those in the first half - also known as the Curse of Depth (Sun et al., 2025, Csord\'as et al., 2025). Using depth-wise analyses, we demonstrate that growth via gradual middle stacking yields more effective utilization of model depth, alters the residual stream structure, and facilitates the formation of permutable computational blocks. In addition, we propose a lightweight modification of MIDAS that yields further improvements in downstream reasoning benchmarks. Overall, this work highlights how the gradual growth of model depth can lead to the formation of distinct computational circuits and overcome the limited depth utilization seen in standard non-grown models.