Research map / Architectures and efficiency / RNNs and normalization
KDP: Simplifying Representation Dynamics in Kernel Space
Zeyu Ma, Wanying Wang, Guchu Zou, Mingang Chen, Jianhong Wu
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 |
|---|---|---|---|
| fold blocks via kernel-space linearization instead of simply removing layers (ablation) | KDP on LLaMA-2-7B · 12 classification benchmarks avg | average accuracy: 34.28 -> 53.11 | Improved |
| fold blocks via kernel-space linearization instead of simply removing layers (ablation) | KDP on LLaMA-2-13B · 12 classification benchmarks avg | average accuracy: 45.01 -> 59.89 | Improved |
| fold blocks via kernel-space linearization instead of simply removing layers (ablation) | KDP on LLaMA-3-8B · 12 classification benchmarks avg | average accuracy: 39.58 -> 52.77 | Improved |
| linearize in learnable RFF kernel space rather than original space with OLS (ablation) | KDP on LLaMA-3-8B · HellaSwag | accuracy: 31.3 -> 61.5 | Improved |
| linearize in learnable RFF kernel space rather than original space with OLS (ablation) | KDP on LLaMA-3-8B · MMLU | accuracy: 26.7 -> 55.5 | Improved |
| replace more consecutive layers in one block (larger Kmax) (ablation) | KDP on LLaMA-2-7B · MMLU | accuracy: 48.5 -> 24.0 | Hurt |
| increase RFF kernel size from 256 to 1024 (ablation) | KDP on LLaMA-2-7B · HellaSwag | accuracy: 33.7 -> 65.1 | Improved |
| add PEFT (output-head LoRA with distillation) after KDP | KDP on LLaMA-3.1-8B · WikiText | perplexity: 205.7 -> 31.7 | Improved |
| add PEFT after KDP | KDP on LLaMA-2-7B · 12 classification benchmarks avg | average accuracy: 53.11 -> 52.52 | No clear effect |
| increase pruning ratio (ablation) | KDP · CMMLU | retained performance | Hurt |
Closest papers by meaning
- DLP: Dynamic Layerwise Pruning in Large Language Models (ICML 2025 · similarity 0.88)
- Layer as Puzzle Pieces: Compressing Large Language Models through Layer Concatenation (NeurIPS 2025 · similarity 0.87)
- AlphaPruning: Using Heavy-Tailed Self Regularization Theory for Improved Layer-wise Pruning of Large Language Models (NeurIPS 2024 · similarity 0.87)
- DISP-LLM: Dimension-Independent Structural Pruning for Large Language Models (NeurIPS 2024 · similarity 0.87)
- Probe Pruning: Accelerating LLMs through Dynamic Pruning via Model-Probing (ICLR 2025 · similarity 0.87)
- SkipGPT: Each Token is One of a Kind (ICML 2025 · similarity 0.87)
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.
