Attention Residuals

• URL: http://arxiv.org/abs/2603.15031v1
• Date: Mon, 16 Mar 2026 09:32:21 GMT
• Title: Attention Residuals
• Abstract summary: Residual connections with PreNorm are standard in modern LLMs, yet they accumulate all layer outputs with fixed unit weights.<n>We propose Attention Residuals (AttnRes), which replaces this fixed accumulation with softmax attention over preceding layer outputs.
• Score: 38.59138244826294
• License: http://creativecommons.org/licenses/by-nc-nd/4.0/
• Abstract: Residual connections with PreNorm are standard in modern LLMs, yet they accumulate all layer outputs with fixed unit weights. This uniform aggregation causes uncontrolled hidden-state growth with depth, progressively diluting each layer's contribution. We propose Attention Residuals (AttnRes), which replaces this fixed accumulation with softmax attention over preceding layer outputs, allowing each layer to selectively aggregate earlier representations with learned, input-dependent weights. To address the memory and communication overhead of attending over all preceding layer outputs for large-scale model training, we introduce Block AttnRes, which partitions layers into blocks and attends over block-level representations, reducing the memory footprint while preserving most of the gains of full AttnRes. Combined with cache-based pipeline communication and a two-phase computation strategy, Block AttnRes becomes a practical drop-in replacement for standard residual connections with minimal overhead. Scaling law experiments confirm that the improvement is consistent across model sizes, and ablations validate the benefit of content-dependent depth-wise selection. We further integrate AttnRes into the Kimi Linear architecture (48B total / 3B activated parameters) and pre-train on 1.4T tokens, where AttnRes mitigates PreNorm dilution, yielding more uniform output magnitudes and gradient distribution across depth, and improves downstream performance across all evaluated tasks.

Related papers

• Scaling Adaptive Depth with Norm-Agnostic Residual Networks [9.219329760729629]
  NAG is a norm-agnostic residual architecture that separates magnitude from directional information in the residual stream.<n>NAG introduces only a negligible number of additional parameters and relies on simple operations that are easily kernel-fusible.<n>We show that this architecture outperforms baseline Transformers, with gains that increase substantially as depth grows.
  arXiv  Detail & Related papers  (2026-06-15T02:05:37Z)
• Latent Recurrent Transformer: Architecture Exploration, Training Strategies, and Scaling Behavior [107.2098567818173]
  Latent Recurrent Transformer (LRT) is a lightweight augmentation of autoregressive transformers.<n>LRT reuses a high-level source-layer hidden state from the previous token as recurrent memory for the next token.
  arXiv  Detail & Related papers  (2026-05-26T10:10:26Z)
• Replacement Learning: Training Neural Networks with Fewer Parameters [15.09968642484538]
  Replacement Learning (RepL) is a training-time paradigm that reduces full-depth redundancy by replacing selected blocks.<n>RepL preserves local contextual continuity while avoiding unnecessary full-layer.<n>We show that RepL reduces trainable parameters, GPU memory usage, and training time while matching or surpassing standard end-to-end training.
  arXiv  Detail & Related papers  (2026-05-19T08:34:31Z)
• Enhancing Gradient Inversion Attacks in Federated Learning via Hierarchical Feature Optimization [56.95448807869383]
  Federated Learning (FL) has emerged as a compelling paradigm for privacy-preserving distributed machine learning.<n>Recent studies find that the gradients exchanged in the FL system are also vulnerable to privacy leakage.<n>We propose textbfGradient textbfInversion over textbfFeature textbfDomains (GIFD)
  arXiv  Detail & Related papers  (2026-04-01T14:32:15Z)
• XAttnRes: Cross-Stage Attention Residuals for Medical Image Segmentation [7.618393856935408]
  Cross-Stage Attention Residuals (XAttnRes) is a mechanism that maintains a global feature history pool accumulating both encoder and decoder stage outputs.<n>XAttnRes consistently improves performance across four datasets and three imaging modalities.
  arXiv  Detail & Related papers  (2026-03-28T15:40:57Z)
• LoGeR: Long-Context Geometric Reconstruction with Hybrid Memory [97.14005794889134]
  We present LoGeR, a novel architecture that scales dense 3D reconstruction to extremely long sequences without post-optimization.<n>LoGeR processes video streams in chunks, leveraging strong bidirectional priors for high-fidelity intra-chunk reasoning.<n>This memory architecture enables LoGeR to be trained on sequences of 128 frames, and generalize up to thousands of frames during inference.
  arXiv  Detail & Related papers  (2026-03-03T18:55:37Z)
• ShiftLUT: Spatial Shift Enhanced Look-Up Tables for Efficient Image Restoration [8.845117852325997]
  ShiftLUT is a novel framework that attains the largest receptive field among all LUT-based methods while maintaining high efficiency.<n>Compared to the previous state-of-the-art method TinyLUT, ShiftLUT achieves a 3.8$times$ larger receptive field and improves an average PSNR by over 0.21 dB.
  arXiv  Detail & Related papers  (2026-03-01T04:00:23Z)
• LESA: Learnable LLM Layer Scaling-Up [57.0510934286449]
  Training Large Language Models (LLMs) from scratch requires immense computational resources, making it prohibitively expensive.<n>Model scaling-up offers a promising solution by leveraging the parameters of smaller models to create larger ones.<n>We propose textbfLESA, a novel learnable method for depth scaling-up.
  arXiv  Detail & Related papers  (2025-02-19T14:58:48Z)
• You Do Not Fully Utilize Transformer's Representation Capacity [4.753535328327317]
  Layer-Integrated Memory (LIMe) is a lightweight extension that learns per-head, per-layer routing weights to integrate representations from all previous layers with negligible overhead.<n>LIMe consistently achieves faster convergence, lower perplexity per FLOP, and substantial accuracy improvements on synthetic tasks.
  arXiv  Detail & Related papers  (2025-02-13T12:00:50Z)
• Learning Inverse Laplacian Pyramid for Progressive Depth Completion [18.977393635158048]
  LP-Net is an innovative framework that implements a multi-scale, progressive prediction paradigm based on Laplacian Pyramid decomposition.<n>At the time of submission, LP-Net ranks 1st among all peer-reviewed methods on the official KITTI leaderboard.
  arXiv  Detail & Related papers  (2025-02-11T06:21:42Z)
• Initialization Matters: On the Benign Overfitting of Two-Layer ReLU CNN with Fully Trainable Layers [20.25049261035324]
  We extend the analysis to two-layer ReLU convolutional neural networks (CNNs) with fully trainable layers. Our results show that the scaling of the output layer is crucial to the training dynamics. In both settings, we provide nearly matching upper and lower bounds on the test errors.
  arXiv  Detail & Related papers  (2024-10-24T20:15:45Z)
• SHERL: Synthesizing High Accuracy and Efficient Memory for Resource-Limited Transfer Learning [63.93193829913252]
  We propose an innovative METL strategy called SHERL for resource-limited scenarios. In the early route, intermediate outputs are consolidated via an anti-redundancy operation. In the late route, utilizing minimal late pre-trained layers could alleviate the peak demand on memory overhead.
  arXiv  Detail & Related papers  (2024-07-10T10:22:35Z)
• BiTAT: Neural Network Binarization with Task-dependent Aggregated Transformation [116.26521375592759]
  Quantization aims to transform high-precision weights and activations of a given neural network into low-precision weights/activations for reduced memory usage and computation. Extreme quantization (1-bit weight/1-bit activations) of compactly-designed backbone architectures results in severe performance degeneration. This paper proposes a novel Quantization-Aware Training (QAT) method that can effectively alleviate performance degeneration.
  arXiv  Detail & Related papers  (2022-07-04T13:25:49Z)
• Dual-constrained Deep Semi-Supervised Coupled Factorization Network with Enriched Prior [80.5637175255349]
  We propose a new enriched prior based Dual-constrained Deep Semi-Supervised Coupled Factorization Network, called DS2CF-Net. To ex-tract hidden deep features, DS2CF-Net is modeled as a deep-structure and geometrical structure-constrained neural network. Our network can obtain state-of-the-art performance for representation learning and clustering.
  arXiv  Detail & Related papers  (2020-09-08T13:10:21Z)

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.