Research map / Architectures and efficiency / Attention and transformers
CARE: Covariance-Aware and Rank-Enhanced Decomposition for Enabling Multi-Head Latent Attention
Zhongzhu Zhou, Fengxiang Bie, Ziyan Chen, Zhenyu Zhang, Yibo Yang, Junxiong Wang, Ben Athiwaratkun, Xiaoxia Wu, Shuaiwen Leon Song
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 |
|---|---|---|---|
| replace weight-only SVD (Palu) with covariance-aware factorization | MLA conversion · One-shot, KV rank 512, Llama-3.1-8B (Alpaca) | avg accuracy (%): 39.33 -> 62.33 (CARE-U) | Improved |
| use covariance-aware factorization instead of Palu(SVD) at rank 128 | MLA conversion · One-shot, KV rank 128, Llama-3.1-8B | WikiText2 perplexity: 3046.58 -> 353.74 (CARE-E) | Improved |
| use adjusted-rank allocation instead of uniform rank (CARE-E vs CARE-U) at rank 128 (ablation) | CARE · One-shot, KV rank 128, Llama-3.1-8B | avg accuracy (%): 36.72 -> 38.59 | Improved |
| use adjusted-rank allocation instead of uniform rank at rank 512 (ablation) | CARE · One-shot, KV rank 512, Llama-3.1-8B | avg accuracy (%): 62.33 -> 57.82 | Hurt |
| use adjusted-rank allocation instead of uniform rank at rank 256 (ablation) | CARE · One-shot, KV rank 256, Qwen3-4B | avg accuracy (%): 54.33 -> 51.23 | Hurt |
| covariance-aware init vs SVD-LLM V2 (activation-aware baseline) at rank 512 | CARE · One-shot, KV rank 512, Llama-3.1-8B | avg accuracy (%): 62.21 (SVD-LLM V2) vs 62.33 (CARE-U) | No clear effect |
| replace TransMLA initialization with CARE(E) covariance-aware init, then heal 1B tokens | TransMLA · Healed Llama-3.1-8B, MLA rank 512 | avg accuracy (%): 61.36 -> 62.59 | Improved |
| replace TransMLA initialization with CARE(E) init, heal 3B tokens | TransMLA · Healed Llama-3.1-8B, MLA rank 512 | avg accuracy (%): 61.86 -> 63.27 | Improved |
| add healing fine-tune from 0B to 3B tokens (ablation) | TransMLA + CARE(E) Init · Llama-3.1-8B, MLA rank 512 | avg accuracy (%): 57.04 -> 63.27 | Improved |
| more calibration samples beyond 512 or longer sequences (ablation) | CARE · Llama-3.1-8B one-shot accuracy | accuracy | No clear effect |
| vary shrinkage coefficient alpha (ablation) | CARE · one-shot | accuracy | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- SVD-LLM V2 — SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression (ICLR 2025) · better on 2 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Virtual Head Attention (NeurIPS 2026 · similarity 0.88)
- Coeff-Tuning: A Graph Filter Subspace View for Tuning Attention-Based Large Models (CVPR 2025 · similarity 0.87)
- SAS: Simulated Attention Score (NeurIPS 2025 · similarity 0.86)
- In-Context Linear Regression Demystified: Training Dynamics and Mechanistic Interpretability of Multi-Head Softmax Attention (ICML 2025 · similarity 0.86)
- MHA2MLA-VLM: Enabling DeepSeek’s Economical Multi-Head Latent Attention Across Vision-Language Models (AAAI 2026 · similarity 0.86)
- MoH: Multi-Head Attention as Mixture-of-Head Attention (ICML 2025 · similarity 0.86)
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.
