Growing Transformers: Modular Composition and Layer-wise Expansion on a Frozen Substrate

• URL: http://arxiv.org/abs/2507.07129v1
• Date: Tue, 08 Jul 2025 20:01:15 GMT
• Title: Growing Transformers: Modular Composition and Layer-wise Expansion on a Frozen Substrate
• Authors: A. Bochkov,
• Abstract summary: This paper explores an alternative, constructive approach to model development, built upon the foundation of non-trainable, deterministic input embeddings.<n>We show that specialist models trained on disparate datasets can be merged into a single, more capable Mixture-of-Experts model.<n>We introduce a layer-wise constructive training methodology, where a deep Transformer is "grown" by progressively stacking and training one layer at a time.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: The prevailing paradigm for scaling large language models (LLMs) involves monolithic, end-to-end training, a resource-intensive process that lacks flexibility. This paper explores an alternative, constructive approach to model development, built upon the foundation of non-trainable, deterministic input embeddings. In prior [1], we established that high-level semantic reasoning can emerge in Transformers using frozen embeddings derived from the visual structure of Unicode glyphs. Here, we demonstrate that this fixed representational substrate acts as a universal "docking port," enabling two powerful and efficient scaling paradigms: seamless modular composition and progressive layer-wise growth. First, we show that specialist models trained on disparate datasets (e.g., Russian and Chinese text) can be merged into a single, more capable Mixture-of-Experts (MoE) model, post-training, with zero architectural modification. This is achieved by simply averaging their output logits. The resulting MoE model exhibits immediate performance improvements on reasoning benchmarks like MMLU, surpassing its constituent experts without catastrophic forgetting. Second, we introduce a layer-wise constructive training methodology, where a deep Transformer is "grown" by progressively stacking and training one layer at a time. This method demonstrates stable convergence and a clear correlation between model depth and the emergence of complex reasoning abilities, such as those required for SQuAD. Our findings suggest a paradigm shift from monolithic optimization towards a more biological or constructive model of AI development, where complexity is built incrementally and modules can be composed freely. This opens new avenues for resource-efficient scaling, continual learning, and a more democratized ecosystem for building powerful AI systems. We release all code and models to facilitate further research.

Related papers

• Flexible Language Modeling in Continuous Space with Transformer-based Autoregressive Flows [46.673228292287895]
  We propose a novel framework that employs transformer-based autoregressive normalizing flows to model continuous representations.<n>This approach unlocks substantial flexibility, enabling the construction of models that can capture global bi-directional context.<n>We propose new mixture-based coupling transformations designed to capture complex dependencies within the latent space shaped by discrete data.
  arXiv  Detail & Related papers  (2025-07-01T04:51:25Z)
• Structural Similarity-Inspired Unfolding for Lightweight Image Super-Resolution [88.20464308588889]
  We propose a Structural Similarity-Inspired Unfolding (SSIU) method for efficient image SR.<n>This method is designed through unfolding an SR optimization function constrained by structural similarity.<n>Our model outperforms current state-of-the-art models, boasting lower parameter counts and reduced memory consumption.
  arXiv  Detail & Related papers  (2025-06-13T14:29:40Z)
• DiM-Gesture: Co-Speech Gesture Generation with Adaptive Layer Normalization Mamba-2 framework [2.187990941788468]
  generative model crafted to create highly personalized 3D full-body gestures solely from raw speech audio. Model integrates a Mamba-based fuzzy feature extractor with a non-autoregressive Adaptive Layer Normalization (AdaLN) Mamba-2 diffusion architecture.
  arXiv  Detail & Related papers  (2024-08-01T08:22:47Z)
• Mamba-FSCIL: Dynamic Adaptation with Selective State Space Model for Few-Shot Class-Incremental Learning [115.79349923044663]
  Few-shot class-incremental learning (FSCIL) aims to incrementally learn novel classes from limited examples.<n>Existing methods face a critical dilemma: static architectures rely on a fixed parameter space to learn from data that arrive sequentially, prone to overfitting to the current session.<n>In this study, we explore the potential of Selective State Space Models (SSMs) for FSCIL.
  arXiv  Detail & Related papers  (2024-07-08T17:09:39Z)
• Orchid: Flexible and Data-Dependent Convolution for Sequence Modeling [4.190836962132713]
  This paper introduces Orchid, a novel architecture designed to address the quadratic complexity of traditional attention mechanisms. At the core of this architecture lies a new data-dependent global convolution layer, which contextually adapts its conditioned kernel on input sequence. We evaluate the proposed model across multiple domains, including language modeling and image classification, to highlight its performance and generality.
  arXiv  Detail & Related papers  (2024-02-28T17:36:45Z)
• Scaling Pre-trained Language Models to Deeper via Parameter-efficient Architecture [68.13678918660872]
  We design a more capable parameter-sharing architecture based on matrix product operator (MPO) MPO decomposition can reorganize and factorize the information of a parameter matrix into two parts. Our architecture shares the central tensor across all layers for reducing the model size.
  arXiv  Detail & Related papers  (2023-03-27T02:34:09Z)
• CSformer: Bridging Convolution and Transformer for Compressive Sensing [65.22377493627687]
  This paper proposes a hybrid framework that integrates the advantages of leveraging detailed spatial information from CNN and the global context provided by transformer for enhanced representation learning. The proposed approach is an end-to-end compressive image sensing method, composed of adaptive sampling and recovery. The experimental results demonstrate the effectiveness of the dedicated transformer-based architecture for compressive sensing.
  arXiv  Detail & Related papers  (2021-12-31T04:37:11Z)
• GroupBERT: Enhanced Transformer Architecture with Efficient Grouped Structures [57.46093180685175]
  We demonstrate a set of modifications to the structure of a Transformer layer, producing a more efficient architecture. We add a convolutional module to complement the self-attention module, decoupling the learning of local and global interactions. We apply the resulting architecture to language representation learning and demonstrate its superior performance compared to BERT models of different scales.
  arXiv  Detail & Related papers  (2021-06-10T15:41:53Z)
• Learning Deep-Latent Hierarchies by Stacking Wasserstein Autoencoders [22.54887526392739]
  We propose a novel approach to training models with deep-latent hierarchies based on Optimal Transport. We show that our method enables the generative model to fully leverage its deep-latent hierarchy, avoiding the well known "latent variable collapse" issue of VAEs.
  arXiv  Detail & Related papers  (2020-10-07T15:04:20Z)
• S2RMs: Spatially Structured Recurrent Modules [105.0377129434636]
  We take a step towards exploiting dynamic structure that are capable of simultaneously exploiting both modular andtemporal structures. We find our models to be robust to the number of available views and better capable of generalization to novel tasks without additional training.
  arXiv  Detail & Related papers  (2020-07-13T17:44:30Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.