ABSTRACT:

As Large Language Models (LLMs) are increasingly deployed in high-stakes domains, ensuring safety and alignment is critical. However, current safety paradigms—such as Reinforcement Learning from Human Feedback (RLHF) and output-level prompt filtering—act only as “surface-level” defenses. They fail to address the underlying computational representations that generate harmful outputs, leaving models vulnerable to “jailbreaks” that bypass filters to activate unsafe internal regimes. The Geometry of Trust shifts the focus from reactive output filtering to proactive, representation-level safety. We treat safety not as a linguistic matching problem, but as a geometric constraint on the model’s internal activation trajectories.

We model the latent space as a structured manifold, shifting from the restrictive “linear representation hypothesis” to a framework where harmful concepts are bounded regions. Our approach integrates:
1) Adaptive Geometric Mapping: Utilizing Riemannian and non-Euclidean tools to define boundaries and “closest safe states” on the manifold. By treating concepts as intrinsic geometric subsets rather than simple directions, we formulate steering as constrained motion along geodesics that avoid forbidden regimes.
2) Causal Disentanglement: We address the problem of “concept entanglement,” where steering harmful behaviors often degrades unrelated model capabilities. By grounding interventions in causal disentanglement and minimal-change projections, we ensure that safety edits are surgically precise, modifying only the targeted factors while preserving the model’s core reasoning and linguistic fluency.
3) Risk Control: Integrating geometric signals, such as distance-to-boundary and curvature-adjusted margins, into a conformal risk control framework to determine when steering is trustworthy and when to trigger deterministic fallbacks. To ensure these interventions reflect genuine model mechanisms rather than lexical artifacts, we incorporate state-of-the-art faithfulness and diagnostic evaluations.

The framework provides a rigorous, prescriptive toolkit for AI alignment, targeting failure mechanisms rather than surface patterns. It offers a scalable, inference-time alternative to expensive fine-tuning. While focused on LLM jailbreak robustness, its impact extends to other high-stakes domains, such as medical imaging, where ensuring models remain within safe, clinically plausible regimes is essential for reliable deployment.