LAVOIR: Gaining Defending Information to Artificial Intelligence Models

Artificial intelligence researchers identified an important limit in the “System One” type decisions models, which can answer questions on the text in a single forward pass: these models remain unable when they encounter missing information. The new method called LAVOIR was developed to solve this problem and teaches models to automatically identify which information should be requested.

Basic findings of research:

• Available problem: Systems such as Laya and TypeSafe’s Jev are going on the forecast path when there is no necessary information in the input text (e.g. if two departments are not specified), which leads to error and uncertainty
• LAVOIR solution: Deficiective information (slot) allows the model to understand which information needs by placing the candidates in the entrance
• Information value account: By using Amortized Value of Information method, every missing information evaluates how much the decision quality can contribute -Single switch advantage: It can efficiently collect information without needing additional inquiry tours

This approach increases the reliability of models that provide** in the natural language processing tasks. In complex critical decision-making systems, the model is aware of its uncertainity and the demand for information can significantly improve the accuracy of outputs. This research published in arXiv provides a new foundation for the development of smarter query strategies of artificial intelligence models.