aiqbench analysis
DtR addresses the quadratic complexity bottleneck of standard Transformers by introducing a distillation-based method to efficiently construct hybrid attention models. This approach replaces full-attention layers with linear alternatives through a single pass, eliminating the need for costly retraining or neural architecture search. The technique offers a practical pathway to balance high expressiveness with linear scaling for diverse downstream tasks.
Distill-then-Replace: Efficient Task-Specific Hybrid Attention Model Construction
arXiv:2601.11667v2 Announce Type: replace-cross Abstract: Transformer architectures deliver state-of-the-art accuracy via dense full-attention, but their quadratic time and memory complexity with respect to sequence length limits practical deployment. Linear attention mechanisms offer linear or near-linear scaling yet often incur performance degradation. Hybrid models that integrate full and linear attention layers promise a balance between efficiency and expressiveness, but face two major challenges: training such hybrid models from scratch is computationally expensive, and manually designing the optimal placement of attention types is highly nontrivial. We propose DtR (Distill-then-Replace), which first transfers weights from the pretrained full-attention modules to its linear attention counterparts through blockwise local distillation, and then applies a greedy layer replacement strategy that iteratively substitutes full attention blocks with linear ones while monitoring validation performance on the target task. DtR 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.
Abstract:Transformer architectures deliver state-of-the-art accuracy via dense full-attention, but their quadratic time and memory complexity with respect to sequence length limits practical deployment. Linear attention mechanisms offer linear or near-linear scaling yet often incur performance degradation. Hybrid models that integrate full and linear attention layers promise a balance between efficiency and expressiveness, but face two major challenges: training such hybrid models from scratch is computationally expensive, and manually designing the optimal placement of attention types is highly nontrivial. We propose DtR (Distill-then-Replace), which first transfers weights from the pretrained full-attention modules to its linear attention counterparts through blockwise local distillation, and then applies a greedy layer replacement strategy that iteratively substitutes full attention blocks with linear ones while monitoring validation performance on the target task. DtR 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.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2601.11667 [cs.LG]
(or arXiv:2601.11667v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2601.11667 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 251 words
Related coverage
[AINews] High Taste Testers Ep. 1
Latent Space · Oct 17, 2026
[AINews] The Future of Latent Space
Latent Space · Sep 25, 2026
[AINews] Gemini 4 Argon: GDM’s answer to Astra/Fable, with 1M output
Latent Space · Oct 1, 2026
Can we change an AI model’s architecture without retraining it from scratch?
Amazon Science · Oct 12, 2026