Research map / Architectures and efficiency / RNNs and normalization
Local Linearity of LLMs Enables Activation Steering via Model-Based Linear Optimal Control
Julian Skifstad, Xinyue Annie Yang, Glen Chou
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 |
|---|---|---|---|
| use A-LQR with linear feature setpoint instead of baseline steering methods for toxicity | A-LQR · RealToxicityPrompts | toxic continuations (%): 5.14 -> 0.12 (Llama-3-8B) | Improved |
| replace S-PID feedback with LQR optimal control on the same setpoint (ablation) | LFS setpoint tracking · RealToxicityPrompts | perplexity / toxicity tradeoff | Improved |
| steer with A-LQR for toxicity on small models | A-LQR · RealToxicityPrompts (Gemma-2-2B) | perplexity: 8.95 -> 12.26 | Hurt |
| steer with A-LQR for truthfulness | A-LQR · TruthfulQA (Llama-3-8B) | T*I: 46.22 -> 63.63 | Improved |
| steer with A-LQR for truthfulness on larger models | A-LQR · TruthfulQA (Qwen-2.5-14B) | MMLU (%): 78.80 -> 71.66 | Hurt |
| add ActAdd steering for truthfulness | ActAdd · TruthfulQA (Gemma-2-2B) | informativeness / MMLU: 96.08 -> 86.80 Info; 51.94 -> 35.34 MMLU | Hurt |
| intervene at all token positions (A-LQR+) instead of last token only for jailbreaking (ablation) | A-LQR · AdvBench (Gemma-2-9B-Instruct) | ASR: 0.1442 -> 0.7692 | Improved |
| intervene at all token positions (A-LQR+) instead of last token only for jailbreaking (ablation) | A-LQR · AdvBench (Qwen2.5-14B-Instruct) | ASR: 0.6923 -> 0.9712 | Improved |
| intervene at all token positions for toxicity/truthfulness/concepts (ablation) | A-LQR · toxicity and truthfulness tasks | primary metrics and PPL/Dist | No clear effect |
| A-LQR last-token steering for jailbreaking vs AAS | A-LQR · AdvBench (Llama-3.2-3B-Instruct) | ASR: 0.9135 (AAS) -> 0.7115 (A-LQR) | Hurt |
| reuse offline-computed LQR gains across reachable activations (linear time-varying approximation) | A-LQR · Gemma-2-2B tracking rollouts | tracking error vs bound | Improved |
Closest papers by meaning
- Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior (NeurIPS 2026 · similarity 0.88)
- Latent Space Chain-of-Embedding Enables Output-free LLM Self-Evaluation (ICLR 2025 · similarity 0.86)
- On-the-Fly Adaptive Distillation of Transformer to Dual-State Linear Attention for Long-Context LLM Serving (ICML 2025 · similarity 0.86)
- The Sharp Directions Are Against You: Curvature Analysis of Activation Steering (NeurIPS 2026 · similarity 0.85)
- Matryoshka Pilot: Learning to Drive Black-Box LLMs with LLMs (NeurIPS 2025 · similarity 0.85)
- Training-Free Task Vectors for LLM Behavioral Control (NeurIPS 2026 · similarity 0.85)
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.
