Sync controls dynamic persona use in ai without retraining
SYNC: A Training-Free Inference Controller for Dynamic Layer Selection in Multi-Persona Alignment
Computation and LanguageArtificial Intelligence
Summary
Using multiple personas in AI models at the same time is tricky because each persona works best at different points inside the model. The authors introduce SYNC, a way to pick the best spots for each persona to influence the AI during use, without any extra training. This method looks at how sensitive each part of the AI is to each persona and adjusts on the fly while generating text. Their tests show SYNC improves multi-persona AI output across various configurations and models.
What this means in practice
- •For ai system builders: Control multiple personas in a single AI model efficiently during inference without retraining or costly offline tuning.
- •For interactive media developers: Create AI-driven experiences that seamlessly switch between different characters or styles by dynamically adjusting persona influence per layer.
Authors
Yosef Nuraga Wicaksana, Muhammad Reza Qorib, Fadhil Muhammad, Alfan Farizki Wicaksono
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 \href{https://github.com/yosefnuragaw/sync}{this https URL}.