Research

Paper

AI LLM March 11, 2026

Prism-$Δ$: Differential Subspace Steering for Prompt Highlighting in Large Language Models

Authors

Yuyao Ge, Shenghua Liu, Yiwei Wang, Tianyu Liu, Baolong Bi, Lingrui Mei, Jiayu Yao, Jiafeng Guo, Xueqi Cheng

Abstract

Prompt highlighting steers a large language model to prioritize user-specified text spans during generation. A key challenge is extracting steering directions that capture the difference between relevant and irrelevant contexts, rather than shared structural patterns common to both. We propose PRISM-$Δ$ (Projection-based Relevance-Informed Steering Method), which decomposes the difference between positive and negative cross-covariance matrices to maximize discriminative energy while eliminating shared directions. Each attention head receives a continuous softplus importance weight, letting weak-but-useful heads contribute at reduced strength. The framework extends naturally to Value representations, capturing content-channel signal that Key-only methods leave unused. Across four benchmarks and five models, PRISM-$Δ$ matches or exceeds the best existing method on 19 of 20 configurations, with relative gains up to +10.6%, while halving the fluency cost of steering. PRISM-$Δ$ also scales to long-context retrieval, outperforming the best existing method by up to +4.8% relative gain. PRISM-$Δ$ is compatible with FlashAttention and adds negligible memory overhead.

Metadata

arXiv ID: 2603.10705
Provider: ARXIV
Primary Category: cs.CL
Published: 2026-03-11
Fetched: 2026-03-12 04:21

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.10705v1</id>\n    <title>Prism-$Δ$: Differential Subspace Steering for Prompt Highlighting in Large Language Models</title>\n    <updated>2026-03-11T12:24:45Z</updated>\n    <link href='https://arxiv.org/abs/2603.10705v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.10705v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Prompt highlighting steers a large language model to prioritize user-specified text spans during generation. A key challenge is extracting steering directions that capture the difference between relevant and irrelevant contexts, rather than shared structural patterns common to both. We propose PRISM-$Δ$ (Projection-based Relevance-Informed Steering Method), which decomposes the difference between positive and negative cross-covariance matrices to maximize discriminative energy while eliminating shared directions. Each attention head receives a continuous softplus importance weight, letting weak-but-useful heads contribute at reduced strength. The framework extends naturally to Value representations, capturing content-channel signal that Key-only methods leave unused. Across four benchmarks and five models, PRISM-$Δ$ matches or exceeds the best existing method on 19 of 20 configurations, with relative gains up to +10.6%, while halving the fluency cost of steering. PRISM-$Δ$ also scales to long-context retrieval, outperforming the best existing method by up to +4.8% relative gain. PRISM-$Δ$ is compatible with FlashAttention and adds negligible memory overhead.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CL'/>\n    <published>2026-03-11T12:24:45Z</published>\n    <arxiv:comment>21 pages, 14 figures</arxiv:comment>\n    <arxiv:primary_category term='cs.CL'/>\n    <author>\n      <name>Yuyao Ge</name>\n    </author>\n    <author>\n      <name>Shenghua Liu</name>\n    </author>\n    <author>\n      <name>Yiwei Wang</name>\n    </author>\n    <author>\n      <name>Tianyu Liu</name>\n    </author>\n    <author>\n      <name>Baolong Bi</name>\n    </author>\n    <author>\n      <name>Lingrui Mei</name>\n    </author>\n    <author>\n      <name>Jiayu Yao</name>\n    </author>\n    <author>\n      <name>Jiafeng Guo</name>\n    </author>\n    <author>\n      <name>Xueqi Cheng</name>\n    </author>\n  </entry>"
}