AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Enhancing Instruction Following of LLMs via Activation Steering with Dynamic Rejection

Minjae Kang, Jaehyung Kim

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Task vectors and steering · 3 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
use plausibility-guided dynamic steering instead of fixed steering strength (ablation)DIRECTER · IFEvalaccuracyImproved
rank layers by attention sensitivity (vs reversed order) (ablation)DIRECTER · IFEvalaccuracy: 79.0 -> 81.8Improved
rank layers by attention sensitivity (vs random layers) (ablation)DIRECTER · IFEvalaccuracy: 80.2 -> 81.8Improved
steer the instruction span instead of random prompt tokens (ablation)DIRECTER · IFEvalaccuracy: 79.2 -> 81.8Improved
add plausibility filter as safety gate to PASTA and SpotLightPASTA, SpotLight · IFEvalaccuracyImproved
apply static steering (PASTA*) to GSM8K-FormatLlama-3.1-8B-Instruct · GSM8K-Formattask accuracy: 82.7 -> 62.7Hurt
apply DIRECTER steering to Llama-3.1-8BLlama-3.1-8B-Instruct · IFEvalprompt accuracy: 73.5 -> 78.8Improved
apply DIRECTER to a small model (Llama-3.2-1B)Llama-3.2-1B-Instruct · IFEvalaccuracy: 61.3 -> 61.6No clear effect
apply DIRECTER to Qwen-2.5 3B/7B/14BQwen-2.5-Instruct · IFEvalaccuracy: 14B: 81.6 -> 83.5; 3B: 63.9 -> 67.1Improved
apply DIRECTER vs zero-shot for generation qualityLlama-3.1-8B-Instruct · LLM-judged text qualitytext quality: 4.39 -> 4.40No clear effect
apply DIRECTER vs zero-shot for task fidelityLlama-3.1-8B-Instruct · LLM-judged task fidelitytask fidelity (%): 84.8 -> 91.7Improved
add DIRECTER decoding loopLlama-3.1-8B-Instruct · throughputthroughput: about 16% lower throughputHurt

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