The AI That Taught Itself: USC Researchers Show How Artificial Intelligence Can Learn What It Never Knew – USC Viterbi
Publish Date: 2026-03-09 14:52:00
Source Domain: viterbischool.usc.edu
- Surprising AI Breakthrough: A new study from USC Viterbi School of Engineering suggests AI can improve past its training data with proper methods.
- Feedback Loop Success: Minda Li and Prof. Bhaskar Krishnamachari developed a feedback loop to greatly enhance GPT-5’s coding ability in the obscure language Idris.
- Testing Obscure Language: The researchers chose Idris, a niche programming language with minimal data, to test their method, even though neither had prior experience with it.
- Remarkable Success Rate Improvement: Initial success rate of GPT-5 on Idris coding problems was 39%, which was dramatically boosted to 96% through iterative feedback and error correction.
- Potential Broad Applications: The researchers propose this feedback-based strategy could significantly enhance AI performance across various domains beyond programming.
- Implications for Low-Resource Languages: The method may enable AI to assist in translation for endangered or less-documented languages with minimal training data.
- Future Enhancements: Future work aims to make the AI model ‘smarter’ by retaining learning across different problems and optimizing the iterative process.
