Research map / Large language models / Language, safety and interpretability
Enhancing Instruction Following of LLMs via Activation Steering with Dynamic Rejection
Minjae Kang, Jaehyung Kim
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 plausibility-guided dynamic steering instead of fixed steering strength (ablation) | DIRECTER · IFEval | accuracy | Improved |
| rank layers by attention sensitivity (vs reversed order) (ablation) | DIRECTER · IFEval | accuracy: 79.0 -> 81.8 | Improved |
| rank layers by attention sensitivity (vs random layers) (ablation) | DIRECTER · IFEval | accuracy: 80.2 -> 81.8 | Improved |
| steer the instruction span instead of random prompt tokens (ablation) | DIRECTER · IFEval | accuracy: 79.2 -> 81.8 | Improved |
| add plausibility filter as safety gate to PASTA and SpotLight | PASTA, SpotLight · IFEval | accuracy | Improved |
| apply static steering (PASTA*) to GSM8K-Format | Llama-3.1-8B-Instruct · GSM8K-Format | task accuracy: 82.7 -> 62.7 | Hurt |
| apply DIRECTER steering to Llama-3.1-8B | Llama-3.1-8B-Instruct · IFEval | prompt accuracy: 73.5 -> 78.8 | Improved |
| apply DIRECTER to a small model (Llama-3.2-1B) | Llama-3.2-1B-Instruct · IFEval | accuracy: 61.3 -> 61.6 | No clear effect |
| apply DIRECTER to Qwen-2.5 3B/7B/14B | Qwen-2.5-Instruct · IFEval | accuracy: 14B: 81.6 -> 83.5; 3B: 63.9 -> 67.1 | Improved |
| apply DIRECTER vs zero-shot for generation quality | Llama-3.1-8B-Instruct · LLM-judged text quality | text quality: 4.39 -> 4.40 | No clear effect |
| apply DIRECTER vs zero-shot for task fidelity | Llama-3.1-8B-Instruct · LLM-judged task fidelity | task fidelity (%): 84.8 -> 91.7 | Improved |
| add DIRECTER decoding loop | Llama-3.1-8B-Instruct · throughput | throughput: about 16% lower throughput | Hurt |
Closest papers by meaning
- Improving Instruction-Following in Language Models through Activation Steering (ICLR 2025 · similarity 0.91)
- Memory Inception: Latent-Space KV Cache Manipulation for Steering LLMs (NeurIPS 2026 · similarity 0.89)
- Focus On This, Not That! Steering LLMs with Adaptive Feature Specification (ICML 2025 · similarity 0.89)
- To Steer or Not to Steer? Mechanistic Error Reduction with Abstention for Language Models (ICML 2025 · similarity 0.88)
- Semantics-Adaptive Activation Intervention for LLMs via Dynamic Steering Vectors (ICLR 2025 · similarity 0.88)
- Steering When Necessary: Flexible Steering Large Language Models with Backtracking (NeurIPS 2025 · similarity 0.88)
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.
