LLMSteer: Improving Long-Context LLM Inference by Steering Attention on Reused Contexts
This addresses efficiency and quality issues for users of large language models, representing an incremental improvement over existing attention steering methods.
The paper tackles the problem of long-context understanding and computational cost in large language models by introducing LLMSteer, a fine-tuning-free framework that uses query-independent attention steering, resulting in narrowing the performance gap with baselines by 65.9% and reducing runtime delay by up to 4.8x.
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.