Research map / Architectures and efficiency / Attention and transformers
LoKiFormer: Locality-aware Attention with Decoupled Knowledge Memory for Efficient Large Language Model Pretraining
Qiuwu Chen, Zimo Liu, Yuchen Li, Ying Sun, Yifan Zhang, Zhijie Qiu, Zeng You, Ryan Dong, Simeng Ma, Yaofo Chen, Mingkui Tan
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 |
|---|---|---|---|
| add Local Fusion Attention (LFA) (ablation) | baseline 5B · Training PPL at 10k steps | perplexity: 31.82 -> 30.57 | Improved |
| add Knowledge Memory Module (KMM) on top of LFA (ablation) | baseline+LFA 5B · Training PPL at 10k steps | perplexity: 30.57 -> 28.50 | Improved |
| add LFA (ablation) | baseline 5B · MMLU zero-shot | accuracy: 17.9 -> 21.2 | Improved |
| add KMM (ablation) | baseline 5B · MMLU zero-shot | accuracy: 17.9 -> 22.9 | Improved |
| add both LFA and KMM (ablation) | baseline 5B · MMLU zero-shot | accuracy: 17.9 -> 25.7 | Improved |
| increase LFA kernel size from 2 to 4 (ablation) | LFA 5B · Training PPL | perplexity: 31.06 -> 30.59 | Improved |
| use head-aligned groups g=h instead of g=1 in LFA (ablation) | LFA 5B · Training PPL at 10k steps | perplexity: 31.07 -> 30.49 | Improved |
| use per-dimension groups g=d instead of g=1 in LFA (ablation) | LFA 5B · Training PPL at 10k steps | perplexity: 31.07 -> 31.49 | Hurt |
| increase knowledge fields F from 0 to 32 (ablation) | KMM 5B · Training PPL cc-en | perplexity: 31.62 -> 30.78 | Improved |
| increase knowledge fields F from 32 to 64 (ablation) | KMM 5B · Training PPL cc-en | perplexity: 30.78 -> 30.35 | Improved |
| increase knowledge fields F from 64 to 128 (ablation) | KMM 5B · Training PPL cc-en | perplexity: 30.35 -> 29.88 | Improved |
| add LFA and KMM to the baseline (ablation) | LoKiFormer-7B · Training throughput (4K context) | tokens/second: 975 -> 950 tokens/second | No clear effect |
Closest papers by meaning
- Star Attention: Efficient LLM Inference over Long Sequences (ICML 2025 · similarity 0.89)
- A Training-Free Sub-quadratic Cost Transformer Model Serving Framework with Hierarchically Pruned Attention (ICLR 2025 · similarity 0.88)
- Loki: Low-rank Keys for Efficient Sparse Attention (NeurIPS 2024 · similarity 0.88)
- A Little Goes a Long Way: Efficient Long Context Training and Inference with Partial Contexts (ICLR 2025 · similarity 0.88)
- CateKV: On Sequential Consistency for Long-Context LLM Inference Acceleration (ICML 2025 · similarity 0.88)
- Peripheral Memory for LLMs: Integration of Sequential Memory Banks with Adaptive Querying (ICML 2025 · similarity 0.88)
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.
