AI Paper Atlas

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

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Safety alignment of LLMs · 14 citations (OpenAlex)

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 changedWhereMeasuredEffect
replace P-only steering vector with full PID controller (Mean-AcT sequential to PID-AcT)Mean-AcT on Gemma2-2B · RealToxicityPromptsCLS toxicity (%): 0.68 -> 0.51Improved
replace P-only steering vector with full PID controller (Mean-AcT sequential to PID-AcT)Mean-AcT on Llama3-8B · RealToxicityPromptsCLS toxicity (%): 1.21 -> 0.72Improved
replace P-only steering with PID in 0-shot LLM-judge toxicityMean-AcT on Llama3-8B · RealToxicityPrompts0-shot toxicity (%): 5.09 -> 4.36Improved
apply steering sequentially across layers (Mean-AcT)Mean-AcT on Gemma2-2B · RealToxicityPromptsCLS toxicity (%): 1.12 -> 0.68Improved
add PID instead of P control on perplexityMean-AcT sequential on Gemma2-2B · Wikipedia perplexityPPL: 14.92 -> 15.22Hurt
add PID instead of P control on MMLUMean-AcT sequential on Llama3-8B · MMLU (5-shot)accuracy: 64.22 -> 64.50No clear effect
replace DIM with PID Steering in Angular Steering jailbreakActAdd on Qwen2.5-14B-Instruct · AdvBenchASR (%): 90.38 -> 92.65Improved
replace DIM with PID Steering in Angular Steering jailbreakActAdd on Llama3.1-8B-Instruct · AdvBenchASR (%): 93.26 -> 94.85Improved
replace DIM with PID Steering in Angular Steering jailbreakActAdd on Gemma2-27B-Instruct · AdvBenchASR (%): 74.03 -> 79.80Improved
replace DIM with PID Steering on tinyGSM8kActAdd on Qwen2.5-14B-Instruct · tinyGSM8k strictaccuracy: 90.12 -> 88.96Hurt
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 harmfulnessstability of improvement: K_d in [0.01,0.05]Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy