Open-Source ntkMirror Released to Prevent Hallucinations

Adopted at ICML 2026, this research leverages the phenomenon of Permutation Dispersion, in which the order in which evidence is presented during grounded question answering (QA) changes the model's answer probabilities.

The newly released ntkMirror is an implementation that can be used immediately on local open-weight models without additional training or a separate model. This technology has the following features:

• ISR=1 Gate: Provides a gate that decides to answer or abstain through theoretical computation without tuning, lowering the hallucination rate to the 0.0~0.7% range.
• No Training Required: Uses the existing model's weights as-is, measuring confidence by scoring evidence across multiple orderings.
• High-Performance Kernel: Through a Fused Kernel that batch-processes permutation operations, it delivers 2.6~10x faster speed compared to existing methods.

Test results showed strong hallucination detection performance, recording an AUROC of 0.94 on the SciFact dataset with the Qwen2.5-7B model, with performance tending to improve as model size increased.