RBS-Attention: Enhancing Long-Context Language Models

Researchers have proposed RBS-Attention, a training-free sparse-prefill method, to improve the efficiency of long-context large language model inference.

The method addresses the 'mean dilution' problem in sparse block selection.

It uses two complementary selection branches.

The goal is to reduce the computational cost of prefill in long-context models.

Source: https://arxiv.org/abs/2609.20971