aiqbench analysis
This study systematically evaluates QKV projection sharing variants, demonstrating that the Q-K=V configuration achieves significant KV cache reductions (up to 96.9% with MQA) with minimal accuracy loss. The findings validate that keys and values can occupy similar representational spaces, making this approach highly effective for reducing memory footprint in on-device inference scenarios. By combining projection sharing with existing head-sharing techniques, researchers can optimize transformer efficiency without compromising performance.
Do Transformers Need Three Projections? Systematic Study of QKV Variants
arXiv:2606.04032v2 Announce Type: replace-cross Abstract: Transformers have become the standard solution for various AI tasks, with the query, key, and value (QKV) attention formulation playing a central role. However, the individual contribution of these three projections and the impact of omitting some remain poorly understood. We systematically evaluate three projection sharing constraints: a) Q-K=V (shared key-value), b) Q=K-V (shared query-key), and c) Q=K=V (single projection). The last two variants produce symmetric attention maps; to address this, we also explore asymmetric attention via 2D positional encodings. Through experiments spanning synthetic tasks, vision (MNIST, CIFAR, TinyImageNet, anomaly), and language modeling (300M and 1.2B parameter models on 10B tokens), we discovered that our transformers perform on par or occasionally better than the QKV transformer. In language modeling, Q-K=V projection sharing achieves 50% KV cache reduction with only 3.1% perplexity degradation. Crucially, projection sharing is complementary to head sharing (GQA/MQA): combining Q-K=V with GQA-4 yields 87.5% cache reduction, while Q-K=V + MQA achieves 96.9%, enabling practical on-device inference. We show that Q-K=V preserves quality because keys and values can occupy similar representational spaces and attention operates in a low-rank regime, whereas Q=K-V breaks attention directionality. Our results systematically characterize projection sharing as an underexplored instance of weight tying in attention, with direct, quantifiable inference memory benefits, particularly valuable for edge deployment. The code is publicly available at https://github.com/Brainchip-Inc/Do-Transformers-Need-3-Projections
Title:Do Transformers Need Three Projections? Systematic Study of QKV Variants
Abstract:Transformers have become the standard solution for various AI tasks, with the query, key, and value (QKV) attention formulation playing a central role. However, the individual contribution of these three projections and the impact of omitting some remain poorly understood. We systematically evaluate three projection sharing constraints: a) Q-K=V (shared key-value), b) Q=K-V (shared query-key), and c) Q=K=V (single projection). The last two variants produce symmetric attention maps; to address this, we also explore asymmetric attention via 2D positional encodings. Through experiments spanning synthetic tasks, vision (MNIST, CIFAR, TinyImageNet, anomaly), and language modeling (300M and 1.2B parameter models on 10B tokens), we discovered that our transformers perform on par or occasionally better than the QKV transformer. In language modeling, Q-K=V projection sharing achieves 50% KV cache reduction with only 3.1% perplexity degradation. Crucially, projection sharing is complementary to head sharing (GQA/MQA): combining Q-K=V with GQA-4 yields 87.5% cache reduction, while Q-K=V + MQA achieves 96.9%, enabling practical on-device inference. We show that Q-K=V preserves quality because keys and values can occupy similar representational spaces and attention operates in a low-rank regime, whereas Q=K-V breaks attention directionality. Our results systematically characterize projection sharing as an underexplored instance of weight tying in attention, with direct, quantifiable inference memory benefits, particularly valuable for edge deployment. The code is publicly available at this https URL
Accepted at ICML 2026 (PMLR vol. 306). 26 pages, 12 figures, 16 tables. Code: this https URL
Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Performance (cs.PF)
I.2.6; I.2.7; I.2.10
arXiv:2606.04032 [cs.LG]
(or arXiv:2606.04032v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2606.04032 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 · 352 words
Related coverage
Physics Transformer: Tailoring Transformer for General PDE Prediction
arXiv · Jul 28, 2026
MoARa: Module-Aware Rank Allocation and Structure-Preserving Decomposition for Low-Rank LLM Pre-training
arXiv · Sep 15, 2026
Lie Routing Vision Language Models for View-Robust 3D Understanding
arXiv · Nov 3, 2026
Do Transformers Need Three Projections? Systematic Study of QKV Variants
arXiv · Jun 4, 2026