Toward Stable Value Alignment: SVGT Achieves Consistent Value Guidance with an Independent Module

To address the problem that large language model values are fragile and hard to stabilize in the dynamic residual stream, the paper proposes the Stable Value Guidance Transformer (SVGT): it maintains normative representations in a value space isolated from the backbone via an independent value module, and converts stable signals into learnable Bridge Tokens to guide the generation trajectory. Experiments on multiple backbone models and safety benchmarks show that SVGT typically reduces harmfulness scores by more than 70% while maintaining generation fluency.