Research map / Architectures and efficiency / Attention and transformers
Multi-Head Low-Rank Attention
Songtao Liu, Hongwu Peng, Zhiwei Zhang, Zhengyu Chen, Yue Guo
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 |
|---|---|---|---|
| use MLRA-4 instead of MLA | MLA · FineWeb-Edu-100B pretraining, 7-dataset avg | perplexity: 13.727 -> 13.672 | Improved |
| use MLRA-4 instead of GQA | GQA · 7-dataset avg | perplexity: 14.139 -> 13.672 | Improved |
| use MLRA-4 instead of MLA | MLA · zero-shot common-sense avg | accuracy: 58.75 -> 58.84 | Improved |
| use 4 branches (MLRA-4) instead of 2 (MLRA-2) (ablation) | MLRA · 7-dataset avg | perplexity: 13.804 -> 13.672 | Improved |
| add gating before attention output projection (ablation) | MLRA-4 · 7-dataset avg | perplexity: 13.672 -> 13.621 | Improved |
| add gating before attention output projection (ablation) | GQA · 7-dataset avg | perplexity: 14.139 -> 13.806 | Improved |
| use zero initialization for output projections instead of N(0, 0.02) (ablation) | all attention variants · pretraining loss and perplexity | loss | Improved |
| add variance scaling of latent states (ablation) | MLA, GLA-2, MLRA-2 · pretraining | perplexity | Improved |
| double the number of attention heads at fixed KV cache (ablation) | GQA, MLA, GLA-2 · pretraining | perplexity | No clear effect |
| use MLRA-4 with TP=4 kernel instead of MLA decoding | MLA · long-context decoding latency | speedup: about 2.8x over MLA | Improved |
| use MLRA-4 kernel instead of GQA | GQA · long-context decoding 128K-2M | speedup: 1.05x-1.26x | Improved |
Closest papers by meaning
- Efficient Low Rank Attention for Long-Context Inference in Large Language Models (NeurIPS 2025 · similarity 0.92)
- Virtual Head Attention (NeurIPS 2026 · similarity 0.91)
- Tensor Product Attention Is All You Need (NeurIPS 2025 · similarity 0.91)
- DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads (ICLR 2025 · similarity 0.91)
- Reducing Transformer Key-Value Cache Size with Cross-Layer Attention (NeurIPS 2024 · similarity 0.90)
- CateKV: On Sequential Consistency for Long-Context LLM Inference Acceleration (ICML 2025 · similarity 0.90)
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.
