USC Engineers Breakthrough Reveals GPT-5 Can Master Obscure Coding Languages Through Real Time Feedback Loops

USC Viterbi researchers show GPT-5 can master the rare Idris language by using feedback loops to fix knowledge gaps, reaching a 96% success rate in real time.

By: AXL Media

Published: Mar 11, 2026, 5:54 AM EDT

Source: The information in this article was sourced from University of Southern California

Challenging the Data Dependent Paradigm of Artificial Intelligence

For years, the development of artificial intelligence has been governed by the principle that a model's utility is strictly limited by the volume of its training data. However, new research from the USC Viterbi School of Engineering, set to be presented at IEEE SoutheastCon 2026, suggests that AI can bridge significant knowledge gaps through iterative self correction. Undergraduate researcher Minda Li and Professor Bhaskar Krishnamachari demonstrated that by providing a model with structured feedback on its own mistakes, it can achieve high levels of mastery in subjects it was barely exposed to during its initial training phase. According to Krishnamachari, this signifies a shift where AI tools are finally able to transcend the data they have seen, effectively learning what they never officially knew.

Testing Capabilities via an Obscure Functional Language

The researchers chose to test GPT-5 using Idris, a functional programming language so niche that it possesses only a fraction of the digital footprint seen by industry standards like Python. While Python boasts over 24 million public code repositories, Idris has approximately 2,000, representing a ten thousandfold decrease in available training material. The choice was intentionally obscure, as neither Li nor Krishnamachari knew how to write a single line of Idris code themselves. This created a unique experimental environment where the human researchers were unable to manually fix the AI's errors, forcing the model to rely entirely on the automated feedback system to navigate a language its own guides could not speak.

The Mechanics of the Compiler Feedback Loop

The breakthrough in performance was achieved through what the team calls a compiler feedback loop. Initially, GPT-5 struggled with Idris coding exercises, failing over 60% of the tasks assigned. To remedy this, Li designed a system that captured the precise technical error messages generated by a compiler and fed them back into the AI model. By allowing the model to attempt a single problem up to 20 times, with each attempt informed by the specific failure of the previous one, the success rate surged from 39% to an impressive 96%. This iterative process transformed the AI from a struggling novice into a specialist, proving that structured external logic can unlock latent capabi...