Embedding Trust: Semantic Isotropy Predicts Nonfactuality in Long-Form Text Generation
Checking a long model response claim by claim is expensive and breaks down on open-ended prompts. This ICML 2026 paper proposes a cheaper first-pass signal. Generate several responses, embed them, and measure how widely those embeddings spread across the unit sphere. Greater dispersion reliably signals lower factual consistency. The method needs no labeled data, no fine-tuning, and no hyperparameter selection, and works with open or closed weight embedding models.
.jpg)
Testimonials
“Our enterprise customers demand trust verification before deploying AI in hiring workflows. Vijil helps us ship AI agents in six weeks instead of six months while dramatically lowering compliance costs.”
“By adapting the Google Responsible Generative AI Toolkit to the needs of enterprises in various industries, Vijil provides critical capabilities for AI developers to preserve the privacy, security and safety of custom models downstream with the same rigor that went into their original release.”
