Prism Transformer: Progressive Head Schedules for Hierarchical Attention Processing

• URL: http://arxiv.org/abs/2606.27449v1
• Date: Thu, 25 Jun 2026 18:18:32 GMT
• Title: Prism Transformer: Progressive Head Schedules for Hierarchical Attention Processing
• Abstract summary: Multi-head attention partitions the hidden dimension equally across all heads at every layer, enforcing an identical representational subspace dimension throughout the models depth.<n>We introduce the Prism Transformer, a novel architectural paradigm that replaces the static, uniform head configuration with a progressive head schedule.<n>By monotonically increasing the head count across layers, the Prism Transformer naturally establishes a local-to-global representational hierarchy.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Multi-head attention conventionally partitions the hidden dimension equally across all heads at every layer, enforcing an identical representational subspace dimension (dh = dmodel/h) throughout the models depth. In this work, we identify this uniform allocation as a fundamental structural bottleneck: due to their restricted dimensional space, early-layer heads are unable to faithfully capture complex, high-dimensional contextual patterns. To resolve this, we introduce the Prism Transformer, a novel architectural paradigm that replaces the static, uniform head configuration with a progressive head schedule. By monotonically increasing the head count across layers, the Prism Transformer naturally establishes a local-to-global representational hierarchy: early layers leverage fewer, exceptionally wide heads to capture complex, local compositional patterns, while deep layers deploy many, narrow heads to decompose these patterns into specialized linguistic features. Crucially, this structural shift is parameter-neutral, compute-neutral, and introduces zero training or inference overhead, preserving identical weight matrices and FLOP budgets as the standard Transformer. Across three model scales (124M, 354M, and 757M), the Prism Transformer consistently outperforms uniform baselines, achieving consistent reductions in validation loss alongside consistent gains on downstream zero-shot benchmarks (including PIQA, HellaSwag, ARC-Easy, and WinoGrande). Our findings demonstrate that non-uniform subspace allocation unlocks latent capacity within the standard Transformer budget, enabling more effective use of model capacity.

Related papers

• Tapered Language Models [17.586289013051864]
  Evidence suggests that layers contribute non-uniformly to the final output, with later refining the residual rather than transforming it.<n>Our controlled experiment shows that, under a fixed budget, allocating more capacity to earlier layers and less to later layers improves perplexity over a uniform-width baseline, while the reverse allocation hurts.
  arXiv  Detail & Related papers  (2026-06-22T17:56:25Z)
• Variable-Width Transformers [57.07167443557886]
  We investigate nonuniform capacity allocation across network depth by proposing a $times$-shaped > former architecture.<n>Our design maintains wider early and late layers while narrowing the middle layers, utilizing a parameter-free residual resizing mechanism.<n>By reducing the average layer width, this architecture also requires fewer overall FLOPs and smaller KV cache memory and I/O cost.
  arXiv  Detail & Related papers  (2026-06-16T17:59:03Z)
• BA-T: An Iterative Transformer for Two-View Bundle Adjustment [53.16990349592508]
  We propose BA-T, an iterative Transformer that implements BA-style structured updates as a repeatable layer in implicit token space.<n> experiments demonstrate that BA-T progressively improves pose and reconstruction accuracy across iterations.<n>BA-T provides a compact, efficient, and structural alternative to depth-heavy attention, enabling accurate 3D reconstruction within a lightweight architecture.
  arXiv  Detail & Related papers  (2026-06-02T07:51:14Z)
• SP-MoMamba: Superpixel-driven Mixture of State Space Experts for Efficient Image Super-Resolution [99.56535031145211]
  State space models (SSMs) have emerged as a powerful paradigm for efficient single-image super-resolution (SR)<n>We propose textbfSP-MoMamba, a superpixel-driven mixture of state space experts for content-aware SR.<n>Our core idea is to transform the traditional rigid scanning into a textbfsemantic-level interaction by treating superpixels as fundamental units.
  arXiv  Detail & Related papers  (2026-05-25T14:19:59Z)
• Structured Multidimensional Representation Learning for Large Language Models [0.0]
  Transformer architectures achieve state-of-the-art performance across a wide range of pattern recognition and natural language processing tasks.<n>We introduce a structured spectral factorization of the embedding space based on the L-product for third-order tensors.<n>We prove that the proposed L-Transformer is spectrally equivalent to p parallel Transformers operating on reduced embeddings.
  arXiv  Detail & Related papers  (2026-03-05T22:34:45Z)
• Hybrid Dual-Path Linear Transformations for Efficient Transformer Architectures [0.0]
  We introduce the Hybrid Dual-Path Linear (HDPL) operator, which decomposes the affine transformation into two topologically distinct pathways.<n> Experiments on the FineWeb-Edu dataset demonstrate that the HDPL architecture outperforms a standard Llama-style baseline.<n>We discuss how the explicit materialization of a probabilistic latent space within the Transformer backbone serves as a vital architectural affordance.
  arXiv  Detail & Related papers  (2026-02-05T20:16:10Z)
• Rethinking Vision Transformer Depth via Structural Reparameterization [16.12815682992294]
  We propose a branch-based structural reparameterization technique that operates during the training phase.<n>Our approach leverages parallel branches within transformer blocks that can be systematically consolidated into streamlined single-path models.<n>When applied to ViT-Tiny, the framework successfully reduces the original 12-layer architecture to 6, 4, or as few as 3 layers while maintaining classification accuracy on ImageNet-1K.
  arXiv  Detail & Related papers  (2025-11-24T21:28:55Z)
• LoG3D: Ultra-High-Resolution 3D Shape Modeling via Local-to-Global Partitioning [26.88556500272625]
  We propose a novel 3D variational autoencoder framework built upon unsigned distance fields (UDFs)<n>Our core innovation is a local-to-global architecture that processes the UDF by partitioning it into uniform subvolumes, UBlocks.<n> Experiments demonstrate state-of-the-art performance in both reconstruction accuracy and generative quality, yielding superior surface smoothness and geometric flexibility.
  arXiv  Detail & Related papers  (2025-11-13T07:34:43Z)
• Understanding Transformers for Time Series: Rank Structure, Flow-of-ranks, and Compressibility [90.894232610821]
  We analyze Transformers through the lens of rank structure.<n>We show that time-series embeddings exhibit sharply decaying singular value spectra.<n>We prove that the associated $Q/K/V$ projections admit accurate low-rank approximations.
  arXiv  Detail & Related papers  (2025-10-02T23:56:17Z)
• SGFormer: Spherical Geometry Transformer for 360 Depth Estimation [52.23806040289676]
  Panoramic distortion poses a significant challenge in 360 depth estimation.<n>We propose a spherical geometry transformer, named SGFormer, to address the above issues.<n>We also present a query-based global conditional position embedding to compensate for spatial structure at varying resolutions.
  arXiv  Detail & Related papers  (2024-04-23T12:36:24Z)
• On the Convergence of Encoder-only Shallow Transformers [62.639819460956176]
  We build the global convergence theory of encoder-only shallow Transformers under a realistic setting. Our results can pave the way for a better understanding of modern Transformers, particularly on training dynamics.
  arXiv  Detail & Related papers  (2023-11-02T20:03:05Z)

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.