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

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, Csordás 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.

Conditionally Invariant Representation Learning for Disentangling Cellular Heterogeneity

This paper presents a novel approach that leverages domain variability to learn representations that are conditionally invariant to unwanted variability or distractors. Our approach identifies both spurious and invariant latent features necessary for achieving accurate reconstruction by placing distinct conditional priors on latent features. The invariant signals are disentangled from noise by enforcing independence which facilitates the construction of an interpretable model with a causal semantic. By exploiting the interplay between data domains and labels, our method simultaneously identifies invariant features and builds invariant predictors. We apply our method to grand biological challenges, such as data integration in single-cell genomics with the aim of capturing biological variations across datasets with many samples, obtained from different conditions or multiple laboratories. Our approach allows for the incorporation of specific biological mechanisms, including gene programs, disease states, or treatment conditions into the data integration process, bridging the gap between the theoretical assumptions and real biological applications. Specifically, the proposed approach helps to disentangle biological signals from data biases that are unrelated to the target task or the causal explanation of interest. Through extensive benchmarking using large-scale human hematopoiesis and human lung cancer data, we validate the superiority of our approach over existing methods and demonstrate that it can empower deeper insights into cellular heterogeneity and the identification of disease cell states.