LLMSteer: Improving Long-Context LLM Inference by Steering Attention on Reused Contexts
Fuente:
arXiv
Guardado en:
| Autores principales: | , , , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2024
|
| Materias: | |
| Acceso en línea: | |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866909398086975488 |
|---|---|
| author | Gu, Zhuohan Yao, Jiayi Du, Kuntai Jiang, Junchen |
| author_facet | Gu, Zhuohan Yao, Jiayi Du, Kuntai Jiang, Junchen |
| contents | As large language models (LLMs) show impressive performance on complex tasks, they still struggle with longer contextual understanding and high computational costs. To balance efficiency and quality, we introduce LLMSteer, a fine-tuning-free framework that enhances LLMs through query-independent attention steering. Tested on popular LLMs and datasets, LLMSteer narrows the performance gap with baselines by 65.9% and reduces the runtime delay by up to 4.8x compared to recent attention steering methods. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2411_13009 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | LLMSteer: Improving Long-Context LLM Inference by Steering Attention on Reused Contexts Gu, Zhuohan Yao, Jiayi Du, Kuntai Jiang, Junchen Machine Learning Computation and Language As large language models (LLMs) show impressive performance on complex tasks, they still struggle with longer contextual understanding and high computational costs. To balance efficiency and quality, we introduce LLMSteer, a fine-tuning-free framework that enhances LLMs through query-independent attention steering. Tested on popular LLMs and datasets, LLMSteer narrows the performance gap with baselines by 65.9% and reduces the runtime delay by up to 4.8x compared to recent attention steering methods. |
| title | LLMSteer: Improving Long-Context LLM Inference by Steering Attention on Reused Contexts |
| topic | Machine Learning Computation and Language |
| url | https://arxiv.org/abs/2411.13009 |