Research map / Architectures and efficiency / Attention and transformers
Spectral Attention Steering for Prompt Highlighting
Weixian Waylon Li, Yuchen Niu, Yongxin Yang, Keshuang Li, Tiejun Ma, Shay B Cohen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| edit key embeddings with SEKA instead of post-hoc attention editing (PASTA) | Qwen3-4B · CounterFact ES | efficacy score: 97.16 -> 99.02 | Improved |
| use SEKA instead of PASTA | Qwen3-14B · Pronoun A. P. Score | all-changed P. score: 90.94 -> 98.54 | Improved |
| use random projections (w/o learn) instead of spectrally learned ones (ablation) | SEKA · CounterFact ES (Qwen3-4B) | efficacy score: 99.02 -> 94.96 | Hurt |
| remove learned projections and KV-head filtering (ablation) | SEKA · Pronoun A. P. Score (Qwen3-4B) | all-changed P. score: 93.26 -> 36.95 | Hurt |
| add query-adaptive expert routing (AdaSEKA) to SEKA | SEKA · Pronoun A. P. Score (Gemma3-4B) | all-changed P. score: 81.11 -> 93.58 | Improved |
| AdaSEKA instead of SEKA | SEKA · CounterFact ES (Gemma3-12B) | efficacy score: 98.86 -> 92.48 | Hurt |
| simple ** marking of highlighted text | Original prompt · Pronoun P. Score (Qwen3-4B) | P. score: 93.14 -> 95.76 | Improved |
| apply SEKA to all passages instead of middle only (ablation) | SEKA · lost-in-the-middle exact match | exact match | Hurt |
| steer middle passages with SEKA | SEKA · lost-in-the-middle exact match | exact match | Improved |
| apply PASTA on lost-in-the-middle | PASTA · lost-in-the-middle exact match | exact match | Hurt |
| use SEKA instead of PASTA for inference | Qwen3-8B-Base · inference time per sample | seconds: PASTA +1.03s vs SEKA +0.03s | Improved |
Closest papers by meaning
- Don't Lose Focus: Activation Steering via Key-Orthogonal Projections (NeurIPS 2026 · similarity 0.85)
- SEA: Sparse Linear Attention with Estimated Attention Mask (ICLR 2024 · similarity 0.85)
- SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking (NeurIPS 2026 · similarity 0.84)
- Coeff-Tuning: A Graph Filter Subspace View for Tuning Attention-Based Large Models (CVPR 2025 · similarity 0.84)
- VIPAMIN: Visual Prompt Initialization via Embedding Selection and Subspace Expansion (NeurIPS 2025 · similarity 0.84)
- Query Efficient Black-Box Visual Prompting with Subspace Learning (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
