Value as Style: Decoupling Value and Semantics with a Unidirectional Mix for Low-Damage LLM Value Steering

The researchers propose a low-damage value steering method that decouples the value orientation of large language models from semantic content, so that activation editing changes normative priorities while preserving the scenarios, facts, and task constraints behind responses. The method builds an editable semantics-value interface on frozen residual states, through a unidirectional semantics-to-value pathway, and learns selective encoding with constraints such as stop-gradient, swap consistency, topic deconfounding, and decorrelation; at inference it fixes the semantic code and edits only the value code to generate residual increments. On two instruction-tuned backbone models, the method improves semantic preservation and reduces benign refusals at a comparable degree of value alignment; on LLaMA-3.1-8B, compared with a prompt method selected on the validation set, alignment is comparable (0.750 vs. 0.748), BERTScore is higher (0.938 vs. 0.923), and contradictions are fewer (5.1% vs. 7.6%). A matched gated-mixture ablation separates the roles of representation learning and selective edit activation, and dimension-matched probes, human ratings, and cross-taxonomy comparisons provide supplementary evidence for low-damage value steering.