Control vectors stay niche in local inference

This feels like a capability that is technically real but still ahead of its deployment curve. The plumbing exists; the missing piece is clear, repeatable operational value.

• –`llama.cpp` supports multiple control vectors, but its own docs warn that stacking too many can degrade output, which makes composition a practical risk rather than a solved feature.
• –`repeng` can export GGUF vectors, so there is a workable train/export/infer path for local models, but the visible use cases still skew toward stylistic steering, sycophancy, and refusal shaping.
• –Dynamic steering is an active research area, with recent work on conditional and context-aware activation steering, but it does not yet look like a common local-inference workflow.
• –The big open problem is vector interaction: nobody seems to have a strong playbook for orthogonality, cancellation, or safe multi-vector blending.
• –If this reaches production, it will likely show up first in narrow, high-value controls like safety policy tuning, persona consistency, and domain-specific tone control, not general-purpose “make the model smarter” hacks.