Modern Transformers Are Implicit Hybrids: From Functional Differentiation to Principled Hybrid Architecture Design

• URL: http://arxiv.org/abs/2609.02986v1
• Date: Wed, 02 Sep 2026 14:49:30 GMT
• Title: Modern Transformers Are Implicit Hybrids: From Functional Differentiation to Principled Hybrid Architecture Design
• Abstract summary: Hybrid architectures combining Full Attention (FA) and Linear Attention (LA) are increasingly prominent, yet their allocation remains.<n>We seek an evidence-grounded basis in head-level functional organization learned by RoPE-based Transformers.<n>We instantiate them in Head-wise Hybrid Architecture (HwH), using NoPE FA for global retrieval and LA for local positional modeling.
• Score: 19.539663057036375
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Hybrid architectures combining Full Attention (FA) and Linear Attention (LA) are increasingly prominent, yet their allocation remains heuristic. We seek an evidence-grounded basis in head-level functional organization learned by RoPE-based Transformers. Behavioral probes do not yield a complete taxonomy, so we propose two intervention metrics: RoPE Frequency Importance Score (RFIS), measuring how each frequency affects a head's attention distribution, and RoPE Positional Dependence (RPD), isolating dependence on rotary positional modulation. On Qwen3-series models and Llama3.1, RFIS suggests and RPD verifies a complete taxonomy of retrieval and positional heads separated by a salient mid-low-frequency band. Controlled Transformers show that this boundary follows the training-length positional scale; we term it the Global Positional Band (GPBand). The analysis suggests a potential cause of zero-shot length-extrapolation failure and yields two principles: positional modeling should operate only locally, with global access through position-independent retrieval; and both functions should be assigned at head granularity with layer-specific allocation. We instantiate them in Head-wise Hybrid Architecture (HwH), using NoPE FA for global retrieval and LA for local positional modeling. With an FA-to-LA ratio below 1:3, HwH retains strong language modeling and commonsense reasoning while improving retrieval and substantially strengthening zero-shot long-context extrapolation over Transformer, LA, and a layer-wise hybrid baseline. Ablations validate both principles and component roles, highlighting principled hybrid architecture design as a promising route toward future foundation models.

Related papers

• Sheaf-Based Federated Representation Learning [11.430396441085579]
  Sheaf-based Federated Representation Learning (SFRL)<n>We develop a decentralized algorithm for solving SFRL, termed Sheaf-FRL, which alternates between gradient updates of the local models and closed-form Procrustes updates of the edge-wise restriction maps.<n>As an application, we consider a cooperative classification task in the context of semantic communication.
  arXiv  Detail & Related papers  (2026-08-08T16:38:34Z)
• Scaling Behavior Foundation Model for Humanoid Robots [77.31115920959364]
  Behavior Foundation Models (BFMs) have emerged as a promising solution to address challenges by leveraging large-scale behavioral data.<n>We show that our approach yields significant improvements in control fidelity and task generalization.<n>These results establish BFM as a principled and effective foundation for scalable and general-purpose humanoid control.
  arXiv  Detail & Related papers  (2026-07-16T16:08:27Z)
• HydraHead: From Head-Level Functional Heterogeneity to Specialized Attention Hybridization [54.02338209487246]
  We introduce HydraHead, a novel architecture that hybridizes Full Attention (FA) and Linear Attention (LA) along the head axis.<n> HydraHead features an interpretability-driven selection strategy that identifies retrieval-critical heads and preserves FA only for them.<n>It achieves over 69% improvement over the baseline at 512K context length, approaching Qwen3.5, a leading model of comparable size with a native context length of 256K.
  arXiv  Detail & Related papers  (2026-06-18T11:14:38Z)
• ReHARK: Refined Hybrid Adaptive RBF Kernels for Robust One-Shot Vision-Language Adaptation [3.8707695363745214]
  ReHARK is a training-free framework that reinterprets few-shot adaptation through global proximal regularization.<n>A new state-of-the-art for one-shot adaptation is established by ReHARK, which achieves an average accuracy of 65.83%.
  arXiv  Detail & Related papers  (2026-03-12T04:59:09Z)
• Revisiting LRP: Positional Attribution as the Missing Ingredient for Transformer Explainability [53.21677928601684]
  Layer-wise relevance propagation is one of the most promising approaches to explainability in deep learning.<n>We propose specialized theoretically-grounded LRP rules designed to propagate attributions across various positional encoding methods.<n>Our method significantly outperforms the state-of-the-art in both vision and NLP explainability tasks.
  arXiv  Detail & Related papers  (2025-06-02T18:07:55Z)
• Fourier Test-time Adaptation with Multi-level Consistency for Robust Classification [10.291631977766672]
  We propose a novel approach called Fourier Test-time Adaptation (FTTA) to integrate input and model tuning. FTTA builds a reliable multi-level consistency measurement of paired inputs for achieving self-supervised of prediction. It was extensively validated on three large classification datasets with different modalities and organs.
  arXiv  Detail & Related papers  (2023-06-05T02:29:38Z)
• Global-to-Local Modeling for Video-based 3D Human Pose and Shape Estimation [53.04781510348416]
  Video-based 3D human pose and shape estimations are evaluated by intra-frame accuracy and inter-frame smoothness. We propose to structurally decouple the modeling of long-term and short-term correlations in an end-to-end framework, Global-to-Local Transformer (GLoT) Our GLoT surpasses previous state-of-the-art methods with the lowest model parameters on popular benchmarks, i.e., 3DPW, MPI-INF-3DHP, and Human3.6M.
  arXiv  Detail & Related papers  (2023-03-26T14:57:49Z)
• Focal and Global Spatial-Temporal Transformer for Skeleton-based Action Recognition [34.38874828210301]
  We propose a novel Focal and Global Spatial-Temporal Transformer network (FG-STFormer) It is equipped with two key components: (1) FG-SFormer: focal joints and global parts coupling spatial transformer. Dilated temporal convolution is integrated into the global self-attention mechanism to explicitly capture the local temporal motion patterns of joints or body parts.
  arXiv  Detail & Related papers  (2022-10-06T05:57:15Z)
• Your Transformer May Not be as Powerful as You Expect [88.11364619182773]
  We mathematically analyze the power of RPE-based Transformers regarding whether the model is capable of approximating any continuous sequence-to-sequence functions. We present a negative result by showing there exist continuous sequence-to-sequence functions that RPE-based Transformers cannot approximate no matter how deep and wide the neural network is. We develop a novel attention module, called Universal RPE-based (URPE) Attention, which satisfies the conditions.
  arXiv  Detail & Related papers  (2022-05-26T14:51:30Z)
• FAMLP: A Frequency-Aware MLP-Like Architecture For Domain Generalization [73.41395947275473]
  We propose a novel frequency-aware architecture, in which the domain-specific features are filtered out in the transformed frequency domain. Experiments on three benchmarks demonstrate significant performance, outperforming the state-of-the-art methods by a margin of 3%, 4% and 9%, respectively.
  arXiv  Detail & Related papers  (2022-03-24T07:26:29Z)

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.