Research map / Large language models / Language, safety and interpretability
Activation Steering with a Feedback Controller
Dung Viet Nguyen, Yen Nhi Pham, Hieu M. Vu, Lei Zhang, 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 P-only steering vector with full PID controller (Mean-AcT sequential to PID-AcT) | Mean-AcT on Gemma2-2B · RealToxicityPrompts | CLS toxicity (%): 0.68 -> 0.51 | Improved |
| replace P-only steering vector with full PID controller (Mean-AcT sequential to PID-AcT) | Mean-AcT on Llama3-8B · RealToxicityPrompts | CLS toxicity (%): 1.21 -> 0.72 | Improved |
| replace P-only steering with PID in 0-shot LLM-judge toxicity | Mean-AcT on Llama3-8B · RealToxicityPrompts | 0-shot toxicity (%): 5.09 -> 4.36 | Improved |
| apply steering sequentially across layers (Mean-AcT) | Mean-AcT on Gemma2-2B · RealToxicityPrompts | CLS toxicity (%): 1.12 -> 0.68 | Improved |
| add PID instead of P control on perplexity | Mean-AcT sequential on Gemma2-2B · Wikipedia perplexity | PPL: 14.92 -> 15.22 | Hurt |
| add PID instead of P control on MMLU | Mean-AcT sequential on Llama3-8B · MMLU (5-shot) | accuracy: 64.22 -> 64.50 | No clear effect |
| replace DIM with PID Steering in Angular Steering jailbreak | ActAdd on Qwen2.5-14B-Instruct · AdvBench | ASR (%): 90.38 -> 92.65 | Improved |
| replace DIM with PID Steering in Angular Steering jailbreak | ActAdd on Llama3.1-8B-Instruct · AdvBench | ASR (%): 93.26 -> 94.85 | Improved |
| replace DIM with PID Steering in Angular Steering jailbreak | ActAdd on Gemma2-27B-Instruct · AdvBench | ASR (%): 74.03 -> 79.80 | Improved |
| replace DIM with PID Steering on tinyGSM8k | ActAdd on Qwen2.5-14B-Instruct · tinyGSM8k strict | accuracy: 90.12 -> 88.96 | Hurt |
| add integral gain K_i to the controller (ablation) | PID Steering on Gemma-2-9B-it jailbreak · Jailbreak harmfulness (LlamaGuard3, LLM-judge) | harmfulness: K_i in [0.05,0.10] | Improved |
| add small derivative gain K_d at high K_i (ablation) | PID Steering on Gemma-2-9B-it jailbreak · Jailbreak harmfulness | stability of improvement: K_d in [0.01,0.05] | Improved |
Closest papers by meaning
- Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior (NeurIPS 2026 · similarity 0.88)
- Angular Steering: Behavior Control via Rotation in Activation Space (NeurIPS 2025 · similarity 0.87)
- The Sharp Directions Are Against You: Curvature Analysis of Activation Steering (NeurIPS 2026 · similarity 0.87)
- Improving Instruction-Following in Language Models through Activation Steering (ICLR 2025 · similarity 0.87)
- To Steer or Not to Steer? Mechanistic Error Reduction with Abstention for Language Models (ICML 2025 · similarity 0.86)
- Steering Externalities: Benign Activation Steering Unintentionally Increases Jailbreak Risk for Large Language Models (NeurIPS 2026 · similarity 0.86)
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.
