What Time Should Artificial Intelligence Agents Keep Human Fixes?
One of the most basic problems encountered when AI systems work with people is how to respond to user interventions. A new research published in ArXiv offers an innovative method to determine that text-based artificial intelligence agents need to appeal to the time to **ne.
The research team is formulated this problem as grounded correct arbitration and defines four basic options of agents:
- Accept user’s correction
- reject correction
- Review and validate the real world
- Request additional information from speaker
- To decide with limited evidence of observation
The application of the system named GAVA is tested with 162 different model control points in a textual environment called ALFWorld. These extensive experiments are trained with accurate and incorrect intervention of 972 pieces. The system can conduct local reviews, make the agent more reliable decisions in conflicting situations.
This study is a critical step for agents working in physical or virtual environments such as robotic and virtual assistants. Unlike the systems that follow human directives blindly, agents using this method can act more autonomous and responsible against incorrect commands. As a result, the safety and effectiveness of human-pay intelligence collaboration can be significantly increased.