Confidence as a Compass: Regime-Dependent Dual Steering for Safety–Utility Tradeoffs in Language Models

Inference-time activation steering promises fine-grained control over language-model behavior, but safety-oriented steering often damages benign-task utility. We introduce confidence-guided dual steering: a two-axis intervention that combines a semantic-safety direction with a temperature-differential confidence direction. Rather than searching for a single best coefficient, we map the full control landscape over semantic-safety strength $\alpha$ and confidence modulation $\beta$. This landscape reveals a regime-dependent picture. In low-to-moderate semantic-safety regimes, confidence modulation can improve the local safety--utility tradeoff relative to semantic-safety-only steering; for example, at $\alpha=0.40$, adding confidence modulation improves balanced efficiency from 0.28 to 0.32 while increasing HarmfulQA safety from 0.23 to 0.29. But stronger semantic-safety steering is not uniformly better: increasing $\alpha$ from 0.30 to 0.90 raises apparent safety from 0.24 to 1.00 while collapsing GSM8K accuracy from 0.31 to 0.00 and TriviaQA accuracy from 0.63 to 0.04. We also show that refusal rate and actual harmfulness diverge, motivating evaluation beyond surface refusals. These results position confidence not as a calibrated probability for a single generation, but as a representation-derived control signal for identifying where steering remains useful and where it breaks down.