This research introduces post-hoc uncertainty estimation for Vision-Language Models (VLMs), moving beyond deterministic mappings to capture model uncertainty.

• Transforms existing deterministic VLMs into probabilistic models without retraining
• Provides reliable uncertainty metrics when facing novel inputs or domain shifts
• Demonstrates superior performance in out-of-distribution detection
• Enhances security by allowing models to express low confidence in potentially harmful or unfamiliar inputs

For security applications, this approach offers critical safeguards by enabling AI systems to recognize when they might make unreliable predictions - essential for deployment in high-stakes environments.

Post-hoc Probabilistic Vision-Language Models