arXiv cs.CL· Yosef Nuraga Wicaksana, Muhammad Reza Qorib, Fadhil Muhammad, Alfan Farizki Wicaksono·· 11 小时前AI 评分22
SYNC：面向多人设对齐的动态层选择免训练推理控制器
SYNC: A Training-Free Inference Controller for Dynamic Layer Selection in Multi-Persona Alignment
AI 导读
SYNC 是一个免训练的推理控制框架，可兼容已有的 steering vectors，通过余弦相似度轨迹动态估计每层对每个人设的敏感度，在每个解码步为每个人设分配最优注入层。在 2 个模型、2 组人设配置及无监督与监督提取的 steering vectors 上，SYNC 在 28 种人设配置中的 21 种领先，其中 15 种显著领先。代码已公开。
正文 · 原文
Abstract:Steering vectors align a model at inference time without touching its weights, but eliciting several personas at once is a different problem from eliciting one. Prior multi-persona steering approaches predetermine both how persona vectors are combined and the specific layers at which they are injected. Moreover, this static commitment introduces substantial offline search costs to identify suitable target layers, despite the fact that optimal intervention strengths vary dynamically across both decoding steps and distinct personas. We introduce SYNC (SensitivitY-guided iNference Control), a novel training-free inference control framework compatible with pre-existing steering vectors. SYNC dynamically estimates each layer's sensitivity to every persona via cosine similarity trajectories, assigning each persona to its optimal injection layer at every decoding step. Across two models, two persona profiles and steering vectors from both unsupervised and supervised extraction, SYNC leads on 21 of 28 persona configurations, significantly on 15. Code is available at this https URL.
| Comments: | Under review |
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2611.02636 [cs.CL] |
| (or arXiv:2611.02636v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2611.02636 arXiv-issued DOI via DataCite |
Submission history
From: Yosef Wicaksana [view email]
[v1]
Mon, 2 Nov 2026 06:59:50 UTC (142 KB)
[v2]
Tue, 3 Nov 2026 05:45:25 UTC (142 KB)
来源：arXiv cs.CL · arxiv.org