Research map / Large language models / Language, safety and interpretability
Spherical Steering: Geometry-Aware Activation Rotation for Language Models
Zejia You, Chunyuan Deng, Hanjie Chen
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 |
|---|---|---|---|
| replace additive steering (CAA) with norm-preserving rotation | LLaMA-3.1-8B-Instruct · TruthfulQA MC average | accuracy: 40.54 -> 53.17 | Improved |
| replace additive steering (ITI) with rotation | LLaMA-3.1-8B-Instruct · TruthfulQA open-ended | TRUE x INFO: 40.31 -> 54.63 | Improved |
| add ITI additive steering to base model | LLaMA-3.1-8B-Instruct · TruthfulQA open-ended | TRUE x INFO: 48.24 -> 40.31 | Hurt |
| apply Spherical Steering | LLaMA-3.1-8B-Instruct · COPA | accuracy: 83.00 -> 95.00 | Improved |
| apply Spherical Steering | LLaMA-3.1-8B-Instruct · MMLU | accuracy: 60.60 -> 62.05 | Improved |
| add vMF confidence gate to rotation (ablation) | Spherical Steering · TruthfulQA (Qwen2.5-7B-Instruct) at high strength | TRUE x INFO | Improved |
| add vMF confidence gate to rotation (ablation) | Spherical Steering · TruthfulQA (Qwen2.5-7B-Instruct) MC | MC accuracy: peak at alpha 0.7 -> 0.8 | Improved |
| steer 2-3 layers instead of 1 (ablation) | Spherical Steering · TruthfulQA (LLaMA-3.1-8B-Instruct) | INFO: 62.92 -> 92.70 | Improved |
| steer 4 or more layers (ablation) | Spherical Steering · TruthfulQA (LLaMA-3.1-8B-Instruct) | MC1: 47.62 -> 41.37 | Hurt |
| increase contrastive data from 25 to 300 questions (ablation) | Spherical Steering · TruthfulQA (LLaMA-3.1-8B-Instruct) | MC1: 51.52 -> 53.92 | Improved |
| increase rotation strength alpha from 0.6 to 1.0 (ablation) | Spherical Steering · TruthfulQA open-ended (LLaMA-3.1-8B, layer 24) | TRUE x INFO: 53.98 -> 11.37 | Hurt |
| increase rotation strength alpha from 0.4 to 1.0 (ablation) | Spherical Steering · TruthfulQA MC1 (LLaMA-3.1-8B, layer 24) | MC1: 32.44 -> 46.15 | Improved |
Closest papers by meaning
- Steering When Necessary: Flexible Steering Large Language Models with Backtracking (NeurIPS 2025 · similarity 0.88)
- Correctness-Optimized Residual Activation Lens (CORAL): Transferrable and Calibration-Aware Inference-Time Steering (ICML 2026 · similarity 0.88)
- Improving Instruction-Following in Language Models through Activation Steering (ICLR 2025 · similarity 0.88)
- Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior (NeurIPS 2026 · similarity 0.88)
- Analysing the Generalisation and Reliability of Steering Vectors (NeurIPS 2024 · similarity 0.88)
- Conditional Optimal Bridge for Riemannian Activation Steering (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.
