Complementary Attention Head Pruning for Efficient Transformers

• URL: http://arxiv.org/abs/2606.19150v1
• Date: Wed, 17 Jun 2026 14:56:27 GMT
• Title: Complementary Attention Head Pruning for Efficient Transformers
• Abstract summary: In this paper, we introduce CAHP, a novel framework that redefines head selection as a global graph-theoretical problem.<n>Rather than evaluating heads in isolation, CAHP graph-based clustering combined with information-theoretic distance measures to identify and preserve a top diverse subset of attention heads.<n>Our structural analysis shows that CAHP avoids the "proximity bias" of gradient-based pruning methods, which tend to preserve heads mainly in layers close to the output, and instead retains a functionally critical set of attention heads in the model's intermediate layers.
• Score: 2.2991119948183525
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: The remarkable success of Transformer-based models in natural language processing stems from architectural scaling, which leads to a large number of parameters and hinders deployment in resource-constrained environments. While structured pruning offers a pathway to compression, existing state-of-the-art methods often rely on gradient-based importance ranking or stochastic gating, which suffer from instability, structural degeneration, and the need for extensive manual hyperparameter tuning. In this paper, we introduce CAHP (Complementary Attention Head Pruning), a novel post-hoc framework that redefines head selection as a global graph-theoretical problem. Rather than evaluating heads in isolation, CAHP utilizes graph-based clustering combined with information-theoretic distance measures to identify and preserve a topologically diverse subset of complementary attention heads. Without requiring a predefined sparsity level or pruning ratio, the framework automatically determines the number of selected attention heads across layers by identifying a diminishing marginal performance curve, where pruning additional heads leads to a sharp degradation in performance, as determined by the chosen polynomial degree. Extensive evaluations on the SST-5 and MNLI benchmarks, across different Transformer model scales, demonstrate that CAHP consistently outperforms competitive baselines, particularly in high-compression regimes. Furthermore, our structural analysis shows that CAHP avoids the "proximity bias" of gradient-based pruning methods, which tend to preserve heads mainly in layers close to the output, and instead retains a functionally critical set of attention heads in the model's intermediate layers.

Related papers

• Relative Kinetic Utility for Reasoning-Aware Structural Pruning in Large Language Models [0.0]
  We propose a novel framework that elevates discrete pruning to a continuous kinetic integral over the depth manifold of a model.<n>Experiments on Qwen-2.5-7B and LLaMA-3-8B improve performance in the high-sparsity regime around 40%.<n>RKU attains 13.34% accuracy on GSM8K at 40% sparsity, outperforming the strongest baseline.
  arXiv  Detail & Related papers  (2026-05-09T15:47:53Z)
• Isotonic Layer: A Universal Framework for Generic Recommendation Debiasing [2.24699417812702]
  We introduce the Isotonic Layer, a novel, differentiable framework that integrates piecewise linear fitting directly into neural architectures.<n>This ensures model outputs remain logically consistent with critical features such as latent relevance, recency, or quality scores.<n>Our approach utilizes a dual task formulation that decouples the recommendation objective into latent relevance estimation and bias aware calibration.
  arXiv  Detail & Related papers  (2026-02-02T20:44:57Z)
• Explicit Multi-head Attention for Inter-head Interaction in Large Language Models [70.96854312026319]
  Multi-head Explicit Attention (MEA) is a simple yet effective attention variant that explicitly models cross-head interaction.<n>MEA shows strong robustness in pretraining, which allows the use of larger learning rates that lead to faster convergence.<n>This enables a practical key-value cache compression strategy that reduces KV-cache memory usage by 50% with negligible performance loss.
  arXiv  Detail & Related papers  (2026-01-27T13:45:03Z)
• Adapting Multimodal Foundation Models for Few-Shot Learning: A Comprehensive Study on Contrastive Captioners [1.2461503242570642]
  This paper presents a study on adapting the Contrastive Captioners (CoCa) visual backbone for few-shot image classification.<n>We identify an "augmentation divergence": while strong data augmentation degrades the performance of linear probing in low-shot settings, it is essential for stabilizing LoRA fine-tuning.<n>We also demonstrate that hybrid objectives incorporating Supervised Contrastive (SupCon) loss yield consistent performance improvements over standard Cross-Entropy.
  arXiv  Detail & Related papers  (2025-12-14T20:13:21Z)
• PointNSP: Autoregressive 3D Point Cloud Generation with Next-Scale Level-of-Detail Prediction [87.33016661440202]
  Autoregressive point cloud generation has long lagged behind diffusion-based approaches in quality.<n>We propose PointNSP, a coarse-to-fine generative framework that preserves global shape structure at low resolutions.<n> Experiments on ShapeNet show that PointNSP establishes state-of-the-art (SOTA) generation quality for the first time within the autoregressive paradigm.
  arXiv  Detail & Related papers  (2025-10-07T06:31:02Z)
• ERIS: An Energy-Guided Feature Disentanglement Framework for Out-of-Distribution Time Series Classification [51.07970070817353]
  An ideal time series classification (TSC) should be able to capture invariant representations.<n>Current methods are largely unguided, lacking the semantic direction required to isolate truly universal features.<n>We propose an end-to-end Energy-Regularized Information for Shift-Robustness framework to enable guided and reliable feature disentanglement.
  arXiv  Detail & Related papers  (2025-08-19T12:13:41Z)
• Where and How to Perturb: On the Design of Perturbation Guidance in Diffusion and Flow Models [48.385105851598745]
  "HeadHunter" is a systematic framework for iteratively selecting attention heads that align with user-centric objectives.<n>"SoftPAG" is a continuous knob to tune perturbation strength and suppress artifacts.<n>We validate our method on modern large-scale DiT-based text-to-image models.
  arXiv  Detail & Related papers  (2025-06-12T17:59:51Z)
• Model Hemorrhage and the Robustness Limits of Large Language Models [119.46442117681147]
  Large language models (LLMs) demonstrate strong performance across natural language processing tasks, yet undergo significant performance degradation when modified for deployment.<n>We define this phenomenon as model hemorrhage - performance decline caused by parameter alterations and architectural changes.
  arXiv  Detail & Related papers  (2025-03-31T10:16:03Z)
• Soft ascent-descent as a stable and flexible alternative to flooding [6.527016551650139]
  We propose a softened, pointwise mechanism called SoftAD that downweights points on the borderline, limits the effects of outliers, and retains the ascent-descent effect of flooding. We demonstrate how SoftAD can realize classification accuracy competitive with flooding while enjoying a much smaller loss generalization gap and model norm.
  arXiv  Detail & Related papers  (2023-10-16T02:02:56Z)
• Domain Generalization Guided by Gradient Signal to Noise Ratio of Parameters [69.24377241408851]
  Overfitting to the source domain is a common issue in gradient-based training of deep neural networks. We propose to base the selection on gradient-signal-to-noise ratio (GSNR) of network's parameters.
  arXiv  Detail & Related papers  (2023-10-11T10:21:34Z)

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.