Research map / Large language models / Instruction and fine-tuning
Steering Language Models with Weight Arithmetic
Constanza Fierro, Fabien Roger
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 contrastive weight steering instead of activation steering to reduce sycophancy | Qwen2.5-7B-Instruct · TruthfulQA/TriviaQA non-sycophancy | non-sycophancy | Improved |
| use all-layers activation steering instead of single-layer to reduce sycophancy (ablation) | activation steering, Qwen2.5-7B-Instruct · TruthfulQA/TriviaQA non-sycophancy | baseline accuracy | Hurt |
| restrict weight steering fine-tuning to MLP bias terms (ablation) | contrastive weight steering · sycophancy mitigation | non-sycophancy | Hurt |
| make weight steering contrastive (tau+ minus tau-) instead of using tau+ alone (ablation) | weight steering · GCD sycophancy (Qwen2.5-1.5B-Instruct) | GCD capability and sycophancy | Improved |
| apply weight steering to a GCD-fine-tuned model | Qwen2.5-1.5B-Instruct · GCD incorrect-solution split | correctness and disagreement | Improved |
| add non-sycophantic examples to training (Joint) for GCD | Qwen2.5-1.5B-Instruct · GCD incorrect-solution split | non-sycophancy | No clear effect |
| add anti-sycophancy system prompt at inference | GCD-fine-tuned Qwen2.5-1.5B-Instruct · GCD incorrect-solution split | non-sycophancy | No clear effect |
| use weight steering rather than activation steering for evil-MCQA | Qwen2.5-7B-Instruct · World Affecting evil MCQA with TinyMMLU | evil rate before capability degradation | Improved |
| use weight steering rather than activation steering for CoT consistency | Qwen2.5-7B-Instruct · evil CoT | CoT inconsistency rate | Improved |
| restrict weight steering to bias terms for evil steering (ablation) | contrastive weight steering · evil MCQA | evil rate | Improved |
| use refusal-data weight steering to restore refusals after GSM8K fine-tuning | Llama-2-7b-chat · DirectHarm4 and GSM-Danger | safety rate | Improved |
| use evilness data for refusal steering | weight and activation steering, Llama-2-7b-chat · DirectHarm4 and GSM-Danger | safety rate | No clear effect |
Closest papers by meaning
- Focus On This, Not That! Steering LLMs with Adaptive Feature Specification (ICML 2025 · similarity 0.90)
- Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior (NeurIPS 2026 · similarity 0.90)
- The Sharp Directions Are Against You: Curvature Analysis of Activation Steering (NeurIPS 2026 · similarity 0.89)
- Improved Representation Steering for Language Models (NeurIPS 2025 · similarity 0.89)
- Steer2Edit: From Activation Steering to Component-Level Editing (NeurIPS 2026 · similarity 0.89)
- Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization (NeurIPS 2024 · similarity 0.89)
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.
