Spectral Attention Steering for Prompt Highlighting

TL;DR

This paper introduces Spectral Editing Key Amplification (SEKA), a training-free method for attention steering that enhances model focus on specific tokens without high memory costs, outperforming existing techniques.

Contribution

The paper proposes SEKA and its adaptive variant AdaSEKA, enabling efficient, training-free attention steering compatible with memory-efficient implementations.

Findings

SEKA significantly improves prompt highlighting performance.

Both methods reduce latency and memory overhead.

Methods outperform strong baselines on standard benchmarks.

Abstract

Attention steering is an important technique for controlling model focus, enabling capabilities such as prompt highlighting, where the model prioritises user-specified text. However, existing attention steering methods require explicit storage of the full attention matrix, making them incompatible with memory-efficient implementations like FlashAttention. We introduce Spectral Editing Key Amplification (SEKA), a training-free steering method that tackles this by directly editing key embeddings before attention computation. SEKA uses spectral decomposition to steer key embeddings towards latent directions that amplify attention scores for certain tokens. We extend this to Adaptive SEKA (AdaSEKA), a query-adaptive variant that uses a training-free routing mechanism to dynamically combine multiple expert subspaces based on the prompt's semantic intent. Our experiments show both methods significantly outperform strong baselines on standard steering benchmarks while adding much lower latency and memory overhead, in compatibility with optimised attention.