Research map / Large language models / Language, safety and interpretability
Activation Steering for LLM Alignment via a Unified ODE-Based Framework
Hongjue Zhao, Haosen Sun, Jiangtao Kong, Xiaochang Li, Qineng Wang, Liwei Jiang, Qi Zhu, Tarek F. Abdelzaher, Yejin Choi, Manling Li, Huajie Shao
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 multi-step ODE solving instead of one step (ablation) | ODESteer (Falcon-7B) · TruthfulQA | T x I: 40.8 -> 42.2 | Improved |
| use multi-step ODE solving instead of one step (ablation) | ODESteer (Mistral-7B) · TruthfulQA | T x I: 58.1 -> 59.9 | Improved |
| use multi-step ODE solving instead of one step (ablation) | ODESteer (LLaMA3.1-8B) · TruthfulQA | T x I: 62.1 -> 63.2 | Improved |
| use multi-step ODE solving instead of one step (ablation) | ODESteer (Falcon-7B) · UltraFeedback | win rate: 54.0 -> 56.3 | Improved |
| use multi-step ODE solving instead of one step (ablation) | ODESteer (Falcon-7B) · RealToxicityPrompts | toxicity (lower better): 0.199 -> 0.188 | Improved |
| use nonlinear density-ratio features (one-step) instead of linear ITI (ablation) | One-step ODESteer (Mistral-7B) · TruthfulQA | T x I: 46.4 -> 58.1 | Improved |
| use nonlinear features (one-step) instead of linear ITI (ablation) | One-step ODESteer (Mistral-7B) · RealToxicityPrompts | toxicity (lower better): 0.165 -> 0.113 | Improved |
| use nonlinear features (one-step) instead of linear ITI (ablation) | One-step ODESteer (Falcon-7B) · UltraFeedback | win rate: 50.5 -> 54.0 | Improved |
| use RK4 solver instead of Euler (ablation) | ODESteer (LLaMA3.1-8B) · TruthfulQA | T x I: 63.2 -> 63.3 | No clear effect |
| use RK4 solver instead of Euler (ablation) | ODESteer (Falcon-7B) · TruthfulQA | T x I: 42.2 -> 42.8 | No clear effect |
| apply ODESteer trained on TruthfulQA to MMLU | ODESteer (LLaMA3.1-8B) · MMLU | accuracy: 61.8 -> 60.9 | No clear effect |
| apply ODESteer trained on TruthfulQA to CommonsenseQA | ODESteer (LLaMA3.1-8B) · CommonsenseQA | accuracy: 68.0 -> 68.3 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TruthFlow — TruthFlow: Truthful LLM Generation via Representation Flow Correction (ICML 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior (NeurIPS 2026 · similarity 0.91)
- Differentially Private Steering for Large Language Model Alignment (ICLR 2025 · similarity 0.90)
- Steering When Necessary: Flexible Steering Large Language Models with Backtracking (NeurIPS 2025 · similarity 0.89)
- Semantics-Adaptive Activation Intervention for LLMs via Dynamic Steering Vectors (ICLR 2025 · similarity 0.89)
- Aligning Large Language Models with Representation Editing: A Control Perspective (NeurIPS 2024 · similarity 0.89)
- ConQuR: Corner Aligned Activation Quantization via Optimized Rotations for LLMs (NeurIPS 2026 · 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.
