Stable Value Guidance Transformer for LLM Alignment

A new paper on arXiv proposes the Stable Value Guidance Transformer (SVGT) to address instability in aligning large language models with human values. The authors identify that values in LLMs are fragile and low-dimensional within the dynamic residual stream, hindering consistent expression. SVGT introduces an independent value module with two designs: independent value modeling, which maintains normative representations in a dedicated space isolated from the backbone, and explicit behavioral guidance, which transduces stable signals into learnable latent Bridge Tokens. These tokens act as dynamic value anchors to steer generative trajectories, ensuring robust adherence across diverse contexts. The paper is available at arXiv:2605.11712.