Code Agents Learn to Generate Harder Math Problems for AI Training

As large language models push toward elite problem-solving standards—competing at levels like the International Mathematical Olympiad—the availability of difficult, diverse math problems has become a critical constraint. Training and evaluating these systems requires a constant supply of fresh, genuinely challenging material, and curating such problems by hand is both costly and slow.

A new study demonstrates that code agents—AI systems trained to write and execute software—can be repurposed as problem generators. Rather than solving math problems, these agents can take existing problems and systematically evolve them into harder variants, creating a scalable pipeline for generating training data.

The Multi-Agent Framework#

Researchers built a collaborative system in which multiple code agents work together. One agent proposes modifications to a math problem, another validates that the new version remains solvable, and a third assesses whether it has genuinely increased in difficulty. By iterating through many candidates, the framework filters out dead ends and keeps only the improvements.

Results and Validation#

Experiments show that with enough computational exploration at test time, code agents produce new problems that are structurally distinct from originals and measurably harder to solve. The generated problems are not mere variations on a theme—they represent genuinely novel problem instances that maintain mathematical coherence.

Broader Implications#

This work suggests that code-driven agents can serve as a viable mechanism for synthesizing difficult mathematical reasoning problems within scalable computational environments. Rather than relying entirely on human mathematicians and problem curators, AI systems themselves can participate in the creation of training data, potentially unlocking faster iteration cycles for advancing machine mathematical reasoning.