Neutrosophic Logic Framework for Epistemic Uncertainty in LLMs

A new study from arXiv (2605.24053v1) proposes Neutrosophic Logic as an alternative to probabilistic frameworks in Large Language Models (LLMs). Traditional LLMs use Softmax layers that constrain probabilities to sum to one, collapsing uncertainty and hindering differentiation between epistemic uncertainty, paradox, and vagueness. The neutrosophic approach treats Truth, Indeterminacy, and Falsity as independent dimensions, allowing their sum to exceed one (hyper-truth). Experiments on four OpenAI GPT models tested five linguistic phenomena—logical paradoxes, epistemic ignorance, vagueness, ethical contradictions, and future contingencies—under three prompting strategies: neutrosophic, probabilistic, and entropy-derived. Results show that neutrosophic prompting better captures epistemic states by permitting T+I+F > 1, offering a new framework for modeling uncertainty in AI.