Research map / Large language models / Instruction and fine-tuning
COLD-Steer: Steering Large Language Models via In-Context One-step Learning Dynamics
Kartik Sharma, Rakshit Trivedi
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 |
|---|---|---|---|
| steer with COLD-FD (finite-difference learning dynamics) instead of unsteered base | COLD-FD · CAA coordinate-ais, Qwen-2.5-7B-Instruct | selection accuracy: 0.02 -> 0.98 | Improved |
| use COLD-FD on Llama-2-7b-chat coordinate-ais | COLD-FD · CAA coordinate-ais, Llama-2-7b-chat | selection accuracy: 0.28 -> 0.90 | Improved |
| use COLD-FD on myopic-reward (Llama-2-7b-chat, pairwise) | COLD-FD · CAA myopic-reward | selection accuracy: 0.76 -> 0.60 | Hurt |
| use COLD-Kernel (unit kernel) instead of base on Qwen | COLD-Kernel · CAA, Qwen-2.5-7B-Instruct | selection accuracy: 0.32 -> 0.32 | No clear effect |
| use COLD-Kernel with positive-only examples on Llama-2-7b-hf | COLD-Kernel · CAA coordinate-ais, Llama-2-7b-hf | selection accuracy: 0.52 -> 0.90 | Improved |
| use COLD-FD on other model families (Gemma, Mistral) | COLD-FD · CAA hallucination | selection accuracy: 0.64 -> 0.70 (Gemma); 0.62 -> 0.88 (Mistral) | Improved |
| use COLD-Kernel for pluralistic alignment instead of base | COLD-Kernel · OpinionsQA, Black respondents | KL divergence: 2.43 -> 0.86 | Improved |
| use COLD-FD for pluralistic alignment | COLD-FD · OpinionsQA, Republicans | KL divergence: 2.38 -> 2.40 | No clear effect |
| use COLD-Kernel instead of ReFT for runtime | COLD-Kernel · CAA pairwise runtime | runtime: 109.57 -> 39.29 | Improved |
| increase number of in-context examples (ablation) | COLD-FD · CAA tasks, Llama-2-7b-hf | selection accuracy | No clear effect |
Closest papers by meaning
- Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior (NeurIPS 2026 · similarity 0.92)
- Focus On This, Not That! Steering LLMs with Adaptive Feature Specification (ICML 2025 · similarity 0.92)
- Steering When Necessary: Flexible Steering Large Language Models with Backtracking (NeurIPS 2025 · similarity 0.91)
- Improved Representation Steering for Language Models (NeurIPS 2025 · similarity 0.91)
- Improving Instruction-Following in Language Models through Activation Steering (ICLR 2025 · similarity 0.90)
- Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization (NeurIPS 2024 · similarity 0.90)
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.
