You Do Not Fully Utilize Transformer's Representation Capacity

• URL: http://arxiv.org/abs/2502.09245v2
• Date: Wed, 28 May 2025 11:04:00 GMT
• Title: You Do Not Fully Utilize Transformer's Representation Capacity
• Abstract summary: 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.
• Score: 4.753535328327317
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: In contrast to RNNs, which compress their history into a single hidden state, Transformers can attend to all past tokens directly. However, standard Transformers rely solely on the hidden state from the previous layer to represent the entire context. We show that this design choice induces representation collapse and degrades performance. To address this issue, we introduce Layer-Integrated Memory (LIMe), a lightweight extension that leverages existing key-value buffers and learns per-head, per-layer routing weights to integrate representations from all previous layers with negligible overhead. Through extensive experiments-including language modeling, synthetic reasoning benchmarks, and very deep architectures-LIMe consistently achieves faster convergence, lower perplexity per FLOP, and substantial accuracy improvements on synthetic tasks while preserving higher value-vector entropy and improved token separability. Finally, our analysis of the learned routing weights reveals systematic reuse of both local and long-distance features, demonstrating how LIMe mitigates collapse, unlocks richer representations without increasing hidden-state size, and points to promising directions for future research.

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)
• 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)
• Attention Residuals [38.59138244826294]
  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.
  arXiv  Detail & Related papers  (2026-03-16T09:32:21Z)
• What matters for Representation Alignment: Global Information or Spatial Structure? [64.67092609921816]
  Representation alignment (REPA) guides generative training by distilling representations from a strong, pretrained vision encoder to intermediate diffusion features.<n>We investigate a fundamental question: what aspect of the target representation matters for generation, its textitglobal revisionsemantic information.<n>We replace the standard projection layer in REPA with a simple convolution layer and introduce a spatial normalization layer for the external representation.
  arXiv  Detail & Related papers  (2025-12-11T16:39:53Z)
• Unleashing Degradation-Carrying Features in Symmetric U-Net: Simpler and Stronger Baselines for All-in-One Image Restoration [52.82397287366076]
  All-in-one image restoration aims to handle diverse degradations (e.g., noise, blur, adverse weather) within a unified framework.<n>In this work, we reveal a critical insight: well-crafted feature extraction inherently encodes degradation-carrying information.<n>Our symmetric design preserves intrinsic degradation signals robustly, rendering simple additive fusion in skip connections.
  arXiv  Detail & Related papers  (2025-12-11T12:20:31Z)
• Mechanisms of Non-Monotonic Scaling in Vision Transformers [0.0]
  We identify a consistent three-phase Cliff-au-Climb pattern that governs how representations evolve with depth.<n>We quantify patterns of information mixing with an Information Scrambling Index, and show that in ViT-L the information-task tradeoff emerges roughly 10 layers later than in ViT-B.
  arXiv  Detail & Related papers  (2025-11-26T18:07:14Z)
• SCOUT: Toward Sub-Quadratic Attention via Segment Compression for Optimized Utility in Transformers [15.142822497807236]
  We propose SCOUT, a hybrid architecture that compresses tokens locally within fixed-size segments and applies attention only over these compressed representations.<n>SCOUT retains much of the expressivity of full attention while substantially reducing the computational and memory cost.<n>We analyze SCOUT's computational and memory efficiency and evaluate it empirically on long-context language modeling and reasoning tasks.
  arXiv  Detail & Related papers  (2025-08-31T17:08:33Z)
• Bottlenecked Transformers: Periodic KV Cache Abstraction for Generalised Reasoning [9.730604030100318]
  Large Language Models struggle with generalisation beyond their training distribution.<n>IB theory posits that model generalisation emerges from an optimal balance between input compression and retention of predictive information in latent representations.<n>We show that decoder-only Transformers are inherently constrained in their ability to form task-optimal sequence representations.<n>We propose a modification to the Transformer architecture, in the form of an additional module that globally rewrites the KV cache.
  arXiv  Detail & Related papers  (2025-05-22T17:33:49Z)
• 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)
• Masked Completion via Structured Diffusion with White-Box Transformers [23.07048591213815]
  We provide the first instantiation of the white-box design paradigm that can be applied to large-scale unsupervised representation learning. We do this by exploiting a fundamental connection between diffusion, compression, and (masked) completion, deriving a deep transformer-like masked autoencoder architecture. CRATE-MAE demonstrates highly promising performance on large-scale imagery datasets.
  arXiv  Detail & Related papers  (2024-04-03T04:23:01Z)
• White-Box Transformers via Sparse Rate Reduction: Compression Is All There Is? [27.58916930770997]
  We show a family of white-box transformer-like deep network architectures, named CRATE, which are mathematically fully interpretable. Experiments show that these networks, despite their simplicity, indeed learn to compress and sparsify representations of large-scale real-world image and text datasets.
  arXiv  Detail & Related papers  (2023-11-22T02:23:32Z)
• 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)
• Jump to Conclusions: Short-Cutting Transformers With Linear Transformations [60.37563766047492]
  Transformer-based language models create hidden representations of their inputs at every layer, but only use final-layer representations for prediction. This obscures the internal decision-making process of the model and the utility of its intermediate representations. We suggest a simple method for such casting, using linear transformations.
  arXiv  Detail & Related papers  (2023-03-16T16:10:16Z)
• Expediting Large-Scale Vision Transformer for Dense Prediction without Fine-tuning [28.180891300826165]
  Many advanced approaches have been developed to reduce the total number of tokens in large-scale vision transformers. We present two non-parametric operators, a token clustering layer to decrease the number of tokens and a token reconstruction layer to increase the number of tokens. Results are promising on five dense prediction tasks, including object detection, semantic segmentation, panoptic segmentation, instance segmentation, and depth estimation.
  arXiv  Detail & Related papers  (2022-10-03T15:49:48Z)
• Vision Transformers: From Semantic Segmentation to Dense Prediction [139.15562023284187]
  We explore the global context learning potentials of vision transformers (ViTs) for dense visual prediction. Our motivation is that through learning global context at full receptive field layer by layer, ViTs may capture stronger long-range dependency information. We formulate a family of Hierarchical Local-Global (HLG) Transformers, characterized by local attention within windows and global-attention across windows in a pyramidal architecture.
  arXiv  Detail & Related papers  (2022-07-19T15:49:35Z)
• Dynamic Spatial Sparsification for Efficient Vision Transformers and Convolutional Neural Networks [88.77951448313486]
  We present a new approach for model acceleration by exploiting spatial sparsity in visual data. We propose a dynamic token sparsification framework to prune redundant tokens. We extend our method to hierarchical models including CNNs and hierarchical vision Transformers.
  arXiv  Detail & Related papers  (2022-07-04T17:00:51Z)
• Plug-In Inversion: Model-Agnostic Inversion for Vision with Data Augmentations [61.95114821573875]
  We introduce Plug-In Inversion, which relies on a simple set of augmentations and does not require excessive hyper- parameter tuning. We illustrate the practicality of our approach by inverting Vision Transformers (ViTs) and Multi-Layer Perceptrons (MLPs) trained on the ImageNet dataset.
  arXiv  Detail & Related papers  (2022-01-31T02:12:45Z)
• Stochastic Layers in Vision Transformers [85.38733795180497]
  We introduce fully layers in vision transformers, without causing any severe drop in performance. The additionality boosts the robustness of visual features and strengthens privacy. We use our features for three different applications, namely, adversarial robustness, network calibration, and feature privacy.
  arXiv  Detail & Related papers  (2021-12-30T16:07:59Z)
• Unifying Global-Local Representations in Salient Object Detection with Transformer [55.23033277636774]
  We introduce a new attention-based encoder, vision transformer, into salient object detection. With the global view in very shallow layers, the transformer encoder preserves more local representations. Our method significantly outperforms other FCN-based and transformer-based methods in five benchmarks.
  arXiv  Detail & Related papers  (2021-08-05T17:51:32Z)
• Visual Saliency Transformer [127.33678448761599]
  We develop a novel unified model based on a pure transformer, Visual Saliency Transformer (VST), for both RGB and RGB-D salient object detection (SOD) It takes image patches as inputs and leverages the transformer to propagate global contexts among image patches. Experimental results show that our model outperforms existing state-of-the-art results on both RGB and RGB-D SOD benchmark datasets.
  arXiv  Detail & Related papers  (2021-04-25T08:24:06Z)
• Vision Transformers for Dense Prediction [77.34726150561087]
  We introduce dense vision transformers, an architecture that leverages vision transformers in place of convolutional networks as a backbone for dense prediction tasks. Our experiments show that this architecture yields substantial improvements on dense prediction tasks.
  arXiv  Detail & Related papers  (2021-03-24T18:01:17Z)
• Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing [112.2208052057002]
  We propose Funnel-Transformer which gradually compresses the sequence of hidden states to a shorter one. With comparable or fewer FLOPs, Funnel-Transformer outperforms the standard Transformer on a wide variety of sequence-level prediction tasks.
  arXiv  Detail & Related papers  (2020-06-05T05:16:23Z)
• Addressing Some Limitations of Transformers with Feedback Memory [51.94640029417114]
  Transformers have been successfully applied to sequential, auto-regressive tasks despite being feedforward networks. We propose the Feedback Transformer architecture that exposes all previous representations to all future representations. We demonstrate on a variety of benchmarks in language modeling, machine translation, and reinforcement learning that the increased representation capacity can create small, shallow models with much stronger performance than comparable Transformers.
  arXiv  Detail & Related papers  (2020-02-21T16:37:57Z)

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.