From Static Problems to Generative Programs: EFAGen Revolutionizes Mathematics

Scientists often derive abstract procedures from specific problem instances and use these abstractions to generate new, related instances. Programs that encode the formal rules and properties of a system have proven useful in fields ranging from reinforcement learning (procedural environments) to physics (simulation engines). These programs can be viewed as functions that lead to different outputs based on their parameterizations (e.g., Gridworld configuration or initial physical conditions). A new approach, called "Executable Functional Abstraction" (EFA), transfers this concept to mathematical problems.

EFAs represent programs that can generate mathematical problems. However, previous work has been limited to abstractions for elementary school-level math problems, whose simple rules can be easily encoded in programs. Generating EFAs for advanced mathematics has previously required human programming. A research team has now investigated the automatic construction of EFAs for more complex mathematical problems.

EFAGen: A New Approach to Automatic EFA Generation

The automatic construction of EFAs is operationalized as a program synthesis task. The developed system, EFAGen, conditions a Large Language Model (LLM) on a mathematical seed problem and its step-by-step solution to generate candidate EFA programs that correspond to the generalized problem and solution class of the seed problem. Furthermore, properties that every valid EFA must possess have been formalized, expressed in executable unit tests. These tests serve as verifiable rewards to train LLMs to write better EFAs.

The results show that EFAs constructed by EFAGen function rationally by staying true to the seed problems and generating learnable problem variations. EFAGen can derive EFAs from various sources of competition-level math problems.

Applications of EFAGen

The EFAs generated by EFAGen offer various applications. They enable the creation of problems with varying difficulty levels for learners. Furthermore, they can be used for data generation to augment training data for AI models and improve their performance.

The automatic generation of EFAs by EFAGen represents a promising approach to revolutionizing mathematics education and research in the field of machine learning. By transforming static mathematical problems into dynamic, generative programs, EFAGen opens new possibilities for personalized learning, automated problem generation, and the exploration of complex mathematical concepts.

Future Research

Research on EFAs and their application is still in its early stages. Future work could focus on improving the efficiency of EFAGen, as well as extending the application areas to other mathematical disciplines. Investigating the impact of EFAs on student learning outcomes is also a promising area of research.

Bibliography:

Suris, D., et al. "ViperGPT: Visual Inference via Python Execution for Reasoning." Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 11471-11481.

Khan, Z., et al. "Executable Functional Abstractions: Inferring Generative Programs for Advanced Math Problems." arXiv preprint arXiv:2504.09763, 2025.

Korinek, A. "Large Language Models (LLMs) and Fintech." Federal Reserve Bank of New York Fintech Conference, 2023.

Calegario, F. "Awesome Generative AI." GitHub repository, https://github.com/filipecalegario/awesome-generative-ai

"DATE 2024 Conference Programme." https://www.date-conference.com/programme

"ICLR 2025 Conference." https://openreview.net/group?id=ICLR.cc/2025/Conference

"Paper QA." GitHub repository, https://github.com/Future-House/paper-qa

Wei, J., et al. "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models." arXiv preprint arXiv:2212.10535, 2022.

Li, Z., et al. "Measuring Faithfulness in Chain-of-Thought Reasoning." arXiv preprint arXiv:2412.16075, 2024.

Brachman, R. J., & Levesque, H. J. (Eds.). Knowledge representation and reasoning. Elsevier, 2004.