USC Researchers Show How Artificial Intelligence Can Learn What It Never Knew

A study from the University of Southern California indicates that artificial intelligence can achieve strong results on tasks outside its primary training data when given appropriate feedback. Researchers Minda Li and Bhaskar Krishnamachari at the USC Viterbi School of Engineering developed and tested the method. The work has been accepted for presentation at IEEE SoutheastCon 2026. Li, an undergraduate, and Krishnamachari, a professor in the Ming Hsieh Department of Electrical and Computer Engineering, focused on the obscure programming language Idris. The language has far fewer code examples available online than widely used languages such as Python. Neither researcher had prior experience writing code in Idris. Initial tests showed GPT-5 solving 22 of 56 exercises on the Exercism platform. This represented a 39 percent success rate. Several methods, including documentation and reference guides, improved results only modestly. The breakthrough occurred when Li created a compiler feedback loop. The system captured compiler error messages and supplied them back to the model for corrections. The model could attempt each problem up to 20 times. Success rose to 96 percent with this approach. Krishnamachari stated that the findings show AI tools can now transcend their initial training. The study suggests the technique could extend to other areas with clear evaluation criteria.