Iteris: Agentic Research Loops for Computational Mathematics

Researchers have developed Iteris, an agentic AI system designed to tackle open problems in computational mathematics by combining language models with numerical experimentation and algorithm design. Applied to two unsolved problems from a Simons Workshop, Iteris generated verified results including a phase diagram for optimization algorithms and a counterexample about QR factorization, demonstrating that AI agents can contribute meaningfully to mathematical research when paired with human expertise.

Iteris represents a meaningful expansion of AI's role in mathematical research beyond traditional theorem-proving domains. While recent advances in large language models have successfully tackled competition mathematics and high-level conjectures, computational mathematics remained underexplored territory—requiring not just symbolic proofs but also numerical evidence, algorithm design, and adversarial constructions. The system's ability to generate verified results on two open problems signals a maturing capability in agentic AI systems to participate in complex research workflows.

This advancement builds on broader trends in AI-assisted discovery. Over the past two years, systems like AlphaProof and similar tools have demonstrated that AI can handle increasingly sophisticated mathematical reasoning. However, Iteris's focus on computational problems—where numerical experimentation and iterative refinement matter as much as logical proof—extends this capability into messier, more applied domains. The emphasis on human review and correction in the published results is particularly notable, as it establishes realistic expectations about AI's role in research rather than claiming autonomy.

For the research and technology communities, Iteris opens pathways for accelerating computational mathematics research and potentially reducing time-to-discovery for applied mathematical problems. This could have downstream applications in optimization, numerical analysis, and scientific computing. The system's architecture suggests a template for agentic AI in other experimental sciences where human-machine collaboration is essential.

Looking forward, the critical questions center on scalability—whether Iteris can tackle genuinely intractable open problems or remains limited to problems amenable to numerical approaches. The field should also monitor whether similar systems emerge in other computational domains and whether human mathematicians' research workflows fundamentally shift as these tools mature.

• →Iteris demonstrates that agentic AI systems can generate verified results on open computational mathematics problems through numerical experimentation and algorithm design.
• →The system successfully produced a phase diagram for optimization algorithm comparison and a counterexample about QR factorization, both validated by expert review.
• →Human expert validation remains essential—AI generated candidate results but did not operate autonomously in the research process.
• →Computational mathematics represents a less-explored frontier for AI compared to pure theorem-proving, offering significant opportunities for human-AI collaboration.
• →This work extends AI's mathematical capabilities from symbolic reasoning into domains requiring numerical evidence and adversarial construction.