Research map / Probabilistic and scientific modeling / Molecules and proteins
Concept Heterogeneity-aware Representation Steering
Laziz Abdullaev, Noelle Y. L. Wong, Ryan Lee T. Z., Shiqi Jiang, Minh-Khoi Nguyen-Nhat, Tan Minh Nguyen
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 |
|---|---|---|---|
| replace ActAdd difference-in-means with CHaRS | ActAdd · AdvBench jailbreak, Gemma2-9B-Instruct | ASR: 91.35 -> 98.08 | Improved |
| replace ActAdd difference-in-means with CHaRS | ActAdd · AdvBench jailbreak, Qwen2.5-32B-Instruct | ASR: 82.69 -> 89.42 | Improved |
| replace DirAbl difference-in-means with CHaRS | DirAbl · AdvBench jailbreak, Gemma2-9B-Instruct | ASR: 79.81 -> 84.62 | Improved |
| replace DirAbl difference-in-means with CHaRS-PCT | DirAbl · AdvBench jailbreak, Gemma2-9B-Instruct | ASR: 79.81 -> 85.58 | Improved |
| replace DirAbl difference-in-means with CHaRS | DirAbl · AdvBench jailbreak, Qwen2.5-14B-Instruct | ASR: 78.85 -> 80.77 | Improved |
| add principal component thresholding to CHaRS | CHaRS with ActAdd · AdvBench jailbreak, Llama3.1-8B-Instruct | ASR: 98.08 -> 99.04 | Improved |
| add principal component thresholding to CHaRS | CHaRS with ActAdd · AdvBench jailbreak, Qwen2.5-7B-Instruct | ASR: 95.19 -> 93.27 | Hurt |
| steer with ActAdd instead of no steering | unsteered Gemma2-9B-Instruct · tinyBenchmarks average | accuracy: 71.11 -> 51.39 | Hurt |
| steer with DirAbl instead of no steering | unsteered Gemma2-9B-Instruct · tinyBenchmarks average | accuracy: 71.11 -> 69.87 | No clear effect |
| use OT coupling instead of nearest-neighbor coupling for cluster matching (ablation) | CHaRS ActAdd · AdvBench jailbreak, Qwen2.5-7B-Instruct, K=5 | ASR: 76.92 -> 95.19 | Improved |
| use OT coupling instead of uniform coupling for cluster matching (ablation) | CHaRS ActAdd · AdvBench jailbreak, Qwen2.5-7B-Instruct, K=5 | ASR: 42.31 -> 95.19 | Improved |
| use more than one cluster (K>1) instead of K=1 (ablation) | CHaRS ActAdd · AdvBench jailbreak, Qwen2.5-7B-Instruct | ASR: 91.35 -> 95.19 | Improved |
Closest papers by meaning
- Improved Representation Steering for Language Models (NeurIPS 2025 · similarity 0.89)
- Representation Surgery: Theory and Practice of Affine Steering (ICML 2024 · similarity 0.89)
- Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization (NeurIPS 2024 · similarity 0.88)
- Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior (NeurIPS 2026 · similarity 0.88)
- Steering When Necessary: Flexible Steering Large Language Models with Backtracking (NeurIPS 2025 · similarity 0.87)
- When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search (NeurIPS 2026 · similarity 0.87)
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.
