Even when a dialogue system produces a content‑safe reply, it may still enforce priorities that clash with the task, e.g., demanding compliance while being asked to preserve autonomy. We formalize this phenomenon as contextual value drift and treat it as a task‑relative decoding problem that explicitly allows value changes.\
LaVGuard separates an authorized ten‑dimensional value profile from a frozen estimate of the values expressed in the visible prefix. A profile‑relative density gate first decides whether to evaluate candidate continuations; then the signed progress of each candidate determines a logit tilt that is constrained by both a changed‑position budget and a cumulative KL budget: $$D_{\text{KL}}(p\|q)\le\epsilon$$.\
We establish local surrogate improvement, pathwise budget feasibility, and a bound on the change of the response distribution. Empirically, conditioning on the profile raises the boundary true‑positive rate at 5% false‑positive rate from 0.58 to 0.72; at the same 0.12 mean intervention ratio, risk‑safety scores are 0.88 for LaVGuard versus 0.72 for position‑matched edits. The resulting interface makes task priorities actionable while keeping the extent of intervention explicit and bounded.\
Review