Researchers' AI agents teach themselves to write code for image-based tasks

A team of researchers from institutions including Georgia Tech and Tencent AI Lab has developed a system where AI models autonomously write and debug their own code to solve complex image-based reasoning tasks. Called self-evolving code-with-image reasoning, the approach eliminates the need for human-crafted algorithms by having models implement visual algorithms directly in Python, using nothing but a code interpreter.

The method works through a training-free reflection loop where models study their own failed attempts, test repairs against ground truth, and retain successful implementations as portable skills. When evaluated on the new Code-with-Image Bench (CwI-Bench) - which contains 30 task families requiring hidden visual computations - the technique dramatically improved performance. For GPT-5.6-luna, accuracy increased from 30% with tool-free chain-of-thought to 43% with just an interpreter, and 67% with evolved executable skills. The 27B open model followed the same progression (9% → 33% → 56%), demonstrating that the evolved skills are text-based and transferable across model scales and task families.