Code2Math's multi-agent framework lets code agents evolve existing math problems into harder, solvable variants under sufficient test-time exploration

The paper introduces a multi-agent framework designed to perform problem evolution while validating the solvability and increased difficulty of the generated problems. Relative to prior practice of relying on human-written or statically collected hard math problems, this places problem generation together with solvability and difficulty checks inside a single multi-agent pipeline. The abstract states the framework is 'designed to perform problem evolution while validating the solvability and increased difficulty of the generated problems,' a method-level design statement.

Experiments show that, given sufficient test-time exploration, code agents can synthesize new, solvable problems that are structurally distinct from and more challenging than the originals. This supplies empirical evidence that code-driven agents can act as a mechanism for synthesizing high-difficulty mathematical reasoning problems, treating code execution as a scalable environment for mathematical experimentation. The abstract reports 'Our experiments demonstrate that, given sufficient test-time exploration' and emphasizes that both solvability and increased difficulty are validated; sample sizes, models, and statistics are not given in the abstract.

The work frames the scarcity of challenging, high-quality problems as a bottleneck for LLM training, evaluation, and self-evolution, and proposes autonomously evolving existing problems with code agents. It shifts the source of problems from human supply toward automatic generation by code agents in scalable computational environments, responding to problem demand as LLM math capabilities advance toward the IMO and research level. The abstract states 'scarcity of challenging, high-quality problems has become a significant bottleneck' as motivation, a problem-positioning statement rather than an experimental measurement.

The framework targets training, evaluation, and self-evolution settings that need high-difficulty mathematical reasoning problems, and applies to computational setups that can provide a code execution environment; its conclusions rest on the condition of 'sufficient test-time exploration,' meaning solvable and harder new problems are observed when the exploration budget is adequate. For researchers and engineers who want to automatically expand math problem sets or study code agents' exploration ability, the work offers a reusable multi-agent pipeline idea.

The abstract does not state which code agents and base models were used, the source and scale of original problems, how difficulty increase was measured, or the specific criteria for solvability validation, nor does it give quantitative results; these details determine the reproducible scope of the conclusions. The abstract also does not discuss whether generated problems might be implicitly isomorphic to originals, or whether difficulty gains hold consistently across mathematical domains, which are open questions for follow-up testing.