Hot eventLive
FTC:序贯结构化压缩LLM注意力无需微调
1 reports1 sources3 hr ago updated
Get the story
AI overview
2026-10-02,arXiv发布一手研究,提出序贯结构化压缩框架FTC。该框架在固定存储预算下适配当前压缩模型,联合利用Q/K/V头结构,并对输出投影单独处理以应对表示变化。实验覆盖6B至32B七个解码器-only LLM,在五款现代GQA模型的WikiText-2困惑度上各keep ratio均最优,激进压缩下增益最大,改进可迁移到下游任务且在32B规模仍显著,无需微调或基于梯度的恢复。
Generated from reports · updated 3 hr ago
Timeline
Follow the coverage from different angles.
Oct 2, 2026
- arXiv · Statistics Machine LearningSequential Functional Structured Tucker Compression for Large Language Model Attentions
FTC是一种序贯结构化压缩框架,在固定存储预算下适配当前压缩模型并联合利用Q/K/V头结构,输出投影单独处理以应对表示变化。实验覆盖6B至32B七个解码器仅LLM,在五款现代GQA模型的WikiText-2困惑度上各keep ratio均最优,激进压缩下增益最大,且改进迁移到下游任务并在32B规模仍显著。无需微调或基于梯度的恢复。
Heat trend
There is not enough continuous observation data to draw a trend yet.