通过旋转激活值实现精准控制，不破坏模型生成能力。
Spherical Steering: Geometry-Aware Activation Rotation for Language Models
- 用球面几何旋转代替加法，保持隐藏状态幅度不变。
- 在TruthfulQA等任务上提升10%，且开放生成质量不受损。
- 适合需要精准控制但不希望破坏模型的场景。
推理时控制为无需微调即可调控语言模型的方法，但传统方法多依赖激活加法，会改变隐藏状态幅值，引发表征坍缩与开放生成质量下降。本文提出球面引导（Spherical Steering），一种无需训练的原始方法，通过沿测地线旋转激活值至目标方向，保持信号完整性的同时实现概念引导。为进一步增强适应性，引入置信度门控机制，根据输入不确定性动态调节引导强度。在多项选择类基准测试中，该方法显著优于基于加法的基线模型（尤其在TruthfulQA、COPA和Storycloze任务上分别提升+10%），同时维持模型的通用开放生成能力。本工作凸显几何一致性的重要性，表明保范旋转是实现精确推理时控制的稳健有效范式。代码已公开：https://github.com/chili-lab/Spherical-Steering。
原文摘要 · Abstract (English)
Inference-time steering offers a promising way to control language models (LMs) without retraining. However, standard approaches typically rely on activation addition, which inevitably alters the hidden-state magnitudes raising concerns about representation collapse and degraded open-ended generation. In this work, we explore Spherical Steering, a training-free primitive that resolves this trade-off through activation rotation. Rather than shifting activations with a fixed vector, our method rotates them along a geodesic toward a target direction, preserving signal integrity while steering toward the target concept. To further enhance adaptivity, we incorporate a confidence gate that dynamically modulates steering strength based on input uncertainty. Extensive experiments across multiple-choice benchmarks demonstrate that Spherical Steering significantly outperforms addition-based baselines (notably by +10% on TruthfulQA, COPA, and Storycloze), while simultaneously maintaining the model's general open-ended generation quality. This work highlights the value of geometric consistency, suggesting that norm-preserving rotation is a robust and effective primitive for precise inference-time control. The code is available at: https://github.com/chili-lab/Spherical-Steering.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。