A local tool aims to cut AI hallucinations by letting models stop
The author says their ICML 2026 paper studies how changing the order of evidence can change a model's answer. They also released ntkMirror, a training-free tool for local open-weight models that decides whether to answer or abstain. In the paper's test, they say the gate reached 0.0% to 0.7% hallucination while abstaining about 24% of the time. They also say a fused kernel made the same gate 2.6 to 10 times faster with little change in accuracy.
Key points
- ntkMirror is released for local open-weight models.
- It checks a claim across several evidence orders before accepting an answer.
- When it judges the information is not enough, it can abstain instead of guessing.
- The author says it needs no fine-tuning and no second model.
- The fused kernel is claimed to speed up the gate without changing accuracy much.
Quick term guide
- training-free
- A method that works without retraining the model.
- open-weight models
- AI models whose internal weights are available for people to run themselves.
- open-weight
- The model's internal numbers are publicly released, so anyone can download and run or modify it freely.
- hallucination
- When AI makes something up and presents it as a real answer.
- AI agents
- AI agents are AI tools that can carry out steps toward a goal, not just answer once.
- agent workflows
- Step-by-step work patterns where an AI agent handles a task.
- agent workflow
- A set of steps an AI follows automatically to complete a series of tasks in order.
- fine-tuning
- Taking an already-trained AI model and doing additional training to specialize it for a specific task.