Attention Editing: A Versatile Framework for Cross-Architecture Attention Conversion

• URL: http://arxiv.org/abs/2604.05688v1
• Date: Tue, 07 Apr 2026 10:40:16 GMT
• Title: Attention Editing: A Versatile Framework for Cross-Architecture Attention Conversion
• Abstract summary: We present Attention Editing, a framework for converting already-trained large language models with new attention architectures without re-pretraining from scratch.<n>We instantiate the framework on two different target--MLA and GateSWA, a gated hybrid SWA design, and apply it to Qwen3-8B and Qwen3-30B-A3B.<n>Experiments are conducted on an Ascend 910B clusters, offering a practical training case study on domestic hardware.
• Score: 17.514757379631927
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Key-Value (KV) cache memory and bandwidth increasingly dominate large language model inference cost in long-context and long-generation regimes. Architectures such as multi-head latent attention (MLA) and hybrid sliding-window attention (SWA) can alleviate this bound, but integrating them into existing models remains difficult. Prior methods impose fine-grained structural requirements on both source and target attention modules, which cannot meet the feasible requirement in practical deployment. We present Attention Editing, a practical framework for converting already-trained large language models (LLMs) with new attention architectures without re-pretraining from scratch. Attention editing replaces the original attention with a learnable target module and trains it using progressive distillation, consisting of (1) layer-wise teacher-forced optimization with intermediate activation supervision to prevent cold-start error accumulation, and (2) model-level distillation on next-token distributions, optionally regularized by weak feature matching. We instantiate the framework on two different target--MLA and GateSWA, a gated hybrid SWA design, and apply it to Qwen3-8B and Qwen3-30B-A3B. The resulting models maintain competitive performance while delivering substantial efficiency improvements, demonstrating that large-scale attention conversion is both feasible and robust. Notably, experiments are conducted on an Ascend 910B clusters, offering a practical training case study on domestic hardware.

Related papers

• MiniCPM-SALA: Hybridizing Sparse and Linear Attention for Efficient Long-Context Modeling [80.48332380100915]
  MiniCPM-SALA is a hybrid model that integrates the high-fidelity long-context modeling of sparse attention with the global efficiency of linear attention.<n>On a single NVIDIA A6000D GPU, the model achieves up to 3.5x the inference speed of the full-attention model at the sequence length of 256K tokens.
  arXiv  Detail & Related papers  (2026-02-12T09:37:05Z)
• Distill-then-Replace: Efficient Task-Specific Hybrid Attention Model Construction [3.9660062354591754]
  Transformer architectures deliver state-of-the-art accuracy via dense full-attention, but their quadratic time and memory complexity limits practical deployment.<n> Linear attention mechanisms offer linear or near-linear scaling yet often incur performance degradation.<n>We introduce a greedy layer replacement strategy that iteratively substitutes full attention blocks with linear ones while monitoring validation performance on the target task.<n>This yields a task-specific hybrid model in a single efficient pass, without costly re-training or neural architecture search, and can be applied to any pretrained full-attention backbone for diverse downstream tasks.
  arXiv  Detail & Related papers  (2026-01-16T02:01:40Z)
• Architecture Decoupling Is Not All You Need For Unified Multimodal Model [64.19284951218098]
  We propose Attention Interaction Alignment (AIA) loss, which explicitly learns Task-Specific multimodal interaction patterns during training.<n>AIA not only refines cross-modal attention patterns, but also boosts both generation and understanding performance.
  arXiv  Detail & Related papers  (2025-11-27T17:55:25Z)
• BOFA: Bridge-Layer Orthogonal Low-Rank Fusion for CLIP-Based Class-Incremental Learning [84.56022893225422]
  Class-Incremental Learning (CIL) aims to continually learn new categories without forgetting previously acquired knowledge.<n>Applying vision-language models such as CLIP to CIL poses two major challenges: (1) adapting to downstream tasks often requires additional learnable modules, increasing model complexity and susceptibility to forgetting; and (2) while multi-modal representations offer complementary strengths, existing methods have yet to fully realize their potential in effectively integrating visual and textual modalities.
  arXiv  Detail & Related papers  (2025-11-14T15:51:40Z)
• DeepThinkVLA: Enhancing Reasoning Capability of Vision-Language-Action Models [51.76664843721462]
  DeepThinkVLA is a new architecture for Vision-Language-Action models.<n>It generates sequential CoT with causal attention and switches to bidirectional attention for fast decoding of action vectors.<n>It achieves a 97.0% success rate on the LIBERO benchmark.
  arXiv  Detail & Related papers  (2025-10-31T05:26:16Z)
• DART: Dual Adaptive Refinement Transfer for Open-Vocabulary Multi-Label Recognition [59.203152078315235]
  Open-Vocabulary Multi-Label Recognition (OV-MLR) aims to identify multiple seen and unseen object categories within an image.<n> Vision-Language Pre-training models offer a strong open-vocabulary foundation, but struggle with fine-grained localization under weak supervision.<n>We propose the Dual Adaptive Refinement Transfer (DART) framework to overcome these limitations.
  arXiv  Detail & Related papers  (2025-08-07T17:22:33Z)
• Cross-Attention Speculative Decoding [20.999323581400336]
  State-of-the-art speculative decoding methods rely on self-attention-based Transformer decoders, often augmented with auxiliary pooling or fusion layers.<n>We present Budget Eagle, a cross-attention-based Transformer decoder SD model that achieves performance on par with leading self-attention SD models.<n>We also propose Two-Stage Block-Attention Training, a new method that achieves training stability and convergence efficiency in block-level attention scenarios.
  arXiv  Detail & Related papers  (2025-05-30T12:52:35Z)
• Scaling Reasoning without Attention [44.42046576158219]
  We introduce ourmodel, an attention-free language model that addresses both issues through architectural and data-centric innovations.<n>Our model eliminates the need for self-attention and key-value caching, enabling fixed-memory, constant-time inference.<n>On benchmark evaluations, ourmodel-7B outperforms strong Transformer and hybrid models of comparable scale.
  arXiv  Detail & Related papers  (2025-05-28T14:52:15Z)
• Hymba: A Hybrid-head Architecture for Small Language Models [65.94140459055244]
  Hymba is a family of small language models featuring a hybrid-head parallel architecture. We introduce learnable meta tokens that are prepended to prompts, storing critical information. This model is further optimized by incorporating cross-layer key-value sharing and partial sliding window attention.
  arXiv  Detail & Related papers  (2024-11-20T19:51:25Z)
• Cross-Architecture Transfer Learning for Linear-Cost Inference Transformers [1.1499643186017316]
  We propose Cross-Architecture Transfer Learning (XATL) to improve efficiency of Transformer Language Models. Methodabbr significantly reduces the training time up to 2.5x times and converges to a better minimum with up to 2.6% stronger model on the LM benchmarks within the same compute budget.
  arXiv  Detail & Related papers  (2024-04-03T12:27:36Z)

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.