Research map / Architectures and efficiency / Attention and transformers
Do Transformers Need Three Projections? Systematic Study of QKV Variants
Ali Kayyam, Anusha Madan Gopal, M Anthony Lewis
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| share key and value projections (K=V) instead of separate V (ablation) | QKV transformer · SlimPajama (300M) | validation perplexity: 5.11 -> 5.27 | Hurt |
| share query and key projections (Q=K) instead of separate Q (ablation) | QKV transformer · SlimPajama (300M) | validation perplexity: 5.11 -> 5.36 | Hurt |
| share a single projection for Q, K and V (ablation) | QKV transformer · SlimPajama (300M) | validation perplexity: 5.11 -> 6.41 | Hurt |
| combine K=V with GQA-4 head sharing (Q-GQA-4) (ablation) | GQA-4 · SlimPajama (300M) | validation perplexity: 5.15 -> 5.32 | Hurt |
| share K=V at 1.2B scale (ablation) | QKV transformer · SlimPajama (1.2B) | perplexity: 5.004 -> 5.128 | Hurt |
| add K=V sharing to Q-MQA at 1.2B (ablation) | MQA · 5-shot downstream average (1.2B) | accuracy (%): 36.37 -> 34.38 | Hurt |
| add K=V sharing to GQA-8 (ablation) | GQA-8 · 5-shot downstream average (1.2B) | accuracy (%): 35.86 -> 36.72 | Improved |
| add 2D positional encoding to Q=K≠V (ablation) | Q=K≠V · Synthetic tasks average | accuracy: 0.724 -> 0.761 | Improved |
| add 2D positional encoding to Q=K=V (ablation) | Q=K=V · Synthetic tasks average | accuracy: 0.651 -> 0.700 | Improved |
| add 2D positional encoding to Q=K≠V (ablation) | Q=K≠V · Vision tasks average | accuracy: 0.766 -> 0.761 | No clear effect |
| share Q=K instead of separate Q (ablation) | QKV transformer · Vision tasks average | accuracy: 0.758 -> 0.766 | No clear effect |
| share Q=K=V instead of separate projections (ablation) | QKV transformer · TinyImageNet | accuracy: 0.331 -> 0.381 | Improved |
Closest papers by meaning
- Understanding Multi-View Transformers (NeurIPS 2026 · similarity 0.85)
- GTA: A Geometry-Aware Attention Mechanism for Multi-View Transformers (ICLR 2024 · similarity 0.85)
- Q Cache: Visual Attention Is Valuable in Less than Half of Decode Layers for Multimodal Large Language Model (AAAI 2026 · similarity 0.85)
- SwitchHead: Accelerating Transformers with Mixture-of-Experts Attention (NeurIPS 2024 · similarity 0.85)
- Mixture-of-Top-$k$ Attention: Efficient Attention as Scalable Fast Weights (NeurIPS 2026 · similarity 0.85)
- No Head Left Behind – Multi-Head Alignment Distillation for Transformers (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
