Residual Stream Duality in Modern Transformer Architectures

• URL: http://arxiv.org/abs/2603.16039v1
• Date: Tue, 17 Mar 2026 00:56:29 GMT
• Title: Residual Stream Duality in Modern Transformer Architectures
• Abstract summary: Recent work has made clear that the residual pathway is not mere optimization plumbing; it is part of the model's representational machinery.<n>We argue that the cleanest way to organize this design space is through a two-axis view of the Transformer.
• Score: 9.910562011343009
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Recent work has made clear that the residual pathway is not mere optimization plumbing; it is part of the model's representational machinery. We agree, but argue that the cleanest way to organize this design space is through a two-axis view of the Transformer. A decoder evolves information along two ordered dimensions: sequence position and layer depth. Self-attention already provides adaptive mixing along the sequence axis, whereas the residual stream usually performs fixed addition along the depth axis. If we fix a token position and treat layer index as the ordered variable, then a causal depth-wise residual attention read is exactly the same local operator as causal short sliding-window attention (ShortSWA), except written over depth rather than over sequence. This is the core residual stream duality behind Transformer$^2$. This perspective also clarifies the recent literature. ELC-BERT and DenseFormer already show that learned aggregation over depth can outperform uniform residual accumulation, while Vertical Attention, DeepCrossAttention (DCA), MUDDFormer, and Attention Residuals move further toward explicit attention-based routing over earlier layers. The key point, however, is that operator-level duality does not imply systems-level symmetry. For large-scale autoregressive models, sequence-axis ShortSWA is usually the more hardware-friendly placement because it reuses token-side sliding-window kernels, KV-cache layouts, and chunked execution. If the goal is instead to change the shortcut itself, Deep Delta Learning (DDL) is the cleaner intervention because it modifies the residual operator directly rather than adding a separate cross-layer retrieval path. Our recommendation is therefore simple: use DDL when the shortcut is the object of interest, and use sequence-axis ShortSWA when the goal is local adaptive mixing.

Related papers

• RASA: Disentangled Spatial-Motional Priors for Cross-Identity Character Animation [85.41886214825894]
  Cross-identity character animation aims to drive a target identity from a reference image to follow the motion of a source character from a driving video.<n>We introduce Reference-Aware Structural Alignment (RASA), a framework that disentangles spatial mapping from motion control.<n>Our work establishes a new paradigm showing disentangled spatial and motional priors are key to robust character animation.
  arXiv  Detail & Related papers  (2026-08-28T11:36:44Z)
• Hidden Decoding at Scale: Latent Computation Scaling for Large Language Models [75.80275843320511]
  We study whether an existing backbone can keep improving by allocating more computation to each token while leaving the Transformer backbone fixed.<n>We propose Hidden Decoding, a sequence-length scaling method applied during continued pretraining.
  arXiv  Detail & Related papers  (2026-07-09T07:37:59Z)
• The Key to Going Linear: Analysis-Driven Transformer Linearization [8.56204304015324]
  We show that softmax relies on key-dependent, rank-1 projections, elucidating why delta-style networks outperform purely gated accumulation.<n>We identify a potential source of approximation errors and introduce structural interventions, specifically sink tokens, short convolutions, and fixed-budget cache routing.
  arXiv  Detail & Related papers  (2026-07-08T17:59:09Z)
• iGSP:Implicit Gradient Subspace Projection for Efficient Continual Learning of Vision-Language Models [14.623213614166259]
  iGSP is a novel framework that achieves efficient adaptation via implicit gradient subspace projection.<n>Experiments on the MTIL benchmark demonstrate that iGSP achieves state-of-the-art accuracy.
  arXiv  Detail & Related papers  (2026-05-19T03:22:14Z)
• Attractor FCM [0.0]
  In this paper an attractor FCM is created, tested, and analyzed.<n>It is gradient descent based, physics constrained, Jacobian version of an FCM.
  arXiv  Detail & Related papers  (2026-04-30T14:44:47Z)
• Stream separation improves Bregman conditioning in transformers [4.7718339202518685]
  Linear methods for steering transformer representations implicitly assume the geometry of representation space is Euclidean.<n>Park et al. showed that softmax induces a curved Bregman geometry whose metric tensor is the Hessian of the log-normalizer, $H() = Cov[ | ]$.<n>We measure this Hessian at intermediate layers in a controlled 2x2 design crossing stream separation with per-layer supervision.
  arXiv  Detail & Related papers  (2026-03-22T16:55:57Z)
• 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)
• Gradient Dynamics of Attention: How Cross-Entropy Sculpts Bayesian Manifolds [0.4779196219827507]
  We show how cross-entropy training reshapes attention scores and value vectors in a transformer attention head.<n>Our core result is an emphadvantage-based routing law for attention scores.<n>We show that this coupled specialization behaves like a two-timescale EM procedure.
  arXiv  Detail & Related papers  (2025-12-27T05:31:44Z)
• SeqPE: Transformer with Sequential Position Encoding [76.22159277300891]
  SeqPE represents each $n$-dimensional position index as a symbolic sequence and employs a lightweight sequential position encoder to learn their embeddings.<n> Experiments across language modeling, long-context question answering, and 2D image classification demonstrate that SeqPE not only surpasses strong baselines in perplexity, exact match (EM) and accuracy--but also enables seamless generalization to multi-dimensional inputs without requiring manual architectural redesign.
  arXiv  Detail & Related papers  (2025-06-16T09:16:40Z)
• Mask-Attention-Free Transformer for 3D Instance Segmentation [68.29828726317723]
  transformer-based methods have dominated 3D instance segmentation, where mask attention is commonly involved. We develop a series of position-aware designs to overcome the low-recall issue and perform cross-attention by imposing positional prior. Experiments show that our approach converges 4x faster than existing work, sets a new state of the art on ScanNetv2 3D instance segmentation benchmark, and also demonstrates superior performance across various datasets.
  arXiv  Detail & Related papers  (2023-09-04T16:09:28Z)
• Attention Map Guided Transformer Pruning for Edge Device [98.42178656762114]
  Vision transformer (ViT) has achieved promising success in both holistic and occluded person re-identification (Re-ID) tasks. We propose a novel attention map guided (AMG) transformer pruning method, which removes both redundant tokens and heads. Comprehensive experiments on Occluded DukeMTMC and Market-1501 demonstrate the effectiveness of our proposals.
  arXiv  Detail & Related papers  (2023-04-04T01:51:53Z)
• URCDC-Depth: Uncertainty Rectified Cross-Distillation with CutFlip for Monocular Depth Estimation [24.03121823263355]
  We introduce an uncertainty rectified cross-distillation between Transformer and convolutional neural network (CNN) to learn a unified depth estimator. Specifically, we use the depth estimates from the Transformer branch and the CNN branch as pseudo labels to teach each other. We propose a surprisingly simple yet highly effective data augmentation technique CutFlip, which enforces the model to exploit more valuable clues apart from the vertical image position for depth inference.
  arXiv  Detail & Related papers  (2023-02-16T08:53:08Z)
• Video Super-Resolution Transformer [85.11270760456826]
  Video super-resolution (VSR), with the aim to restore a high-resolution video from its corresponding low-resolution version, is a spatial-temporal sequence prediction problem. Recently, Transformer has been gaining popularity due to its parallel computing ability for sequence-to-sequence modeling. In this paper, we present a spatial-temporal convolutional self-attention layer with a theoretical understanding to exploit the locality information.
  arXiv  Detail & Related papers  (2021-06-12T20:00:32Z)

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.