Artificial Intelligence for Mathematical Reasoning: An Integrated Survey of Language Models, Neuro-symbolic Systems, and Verified Discovery

A comprehensive survey examines the evolution of AI systems for mathematical reasoning, from early rule-based solvers to contemporary language models, neuro-symbolic systems, and verified discovery workflows. The research catalogs major benchmarks, identifies critical failure modes like reward hacking and formalization brittleness, and proposes future directions centered on efficiency and usable AI-assisted formalization.

This survey represents a significant scholarly effort to map the entire landscape of AI mathematical reasoning, a field that has grown from academic niche to major research frontier. The work progresses chronologically through multiple paradigm shifts—from symbolic rule-based systems through neural approaches to hybrid neuro-symbolic methods—reflecting how machine learning has fundamentally transformed approaches to formal and informal reasoning tasks.

The research addresses a critical gap in AI development: mathematical reasoning remains one of the most rigorous tests of machine intelligence because it requires precision, logical consistency, and often verification against ground truth. The survey's organization across four axes (informal reasoning, formal proving, discovery, and techniques) provides practitioners with a structured taxonomy for understanding where capabilities currently stand and where gaps persist.

For the AI development community, this work has immediate practical implications. The detailed examination of failure modes—particularly reward hacking, brittleness under perturbation, and fragile formalization—highlights engineering challenges that affect real-world deployment. The distinction between pass@1 and verifier-assisted metrics reveals how benchmark reporting can obscure actual performance, a concern for researchers evaluating competing systems. The emphasis on energy costs of reasoning-scale inference touches on operational sustainability, increasingly important as models scale.

The survey's identification of verified-discovery workflows as a future direction suggests the field is moving toward systems that not only solve problems but generate human-verifiable proofs. This shift toward interpretability and formal verification could influence how mathematical AI systems are designed and trusted, with implications extending beyond academia into computational science and automated discovery applications.

• →Mathematical reasoning systems have evolved from symbolic rule-based approaches through neural networks to contemporary neuro-symbolic and multi-agent architectures.
• →Critical failure modes including reward hacking, formalization brittleness, and multimodal grounding errors remain significant engineering challenges.
• →Benchmark saturation and contamination issues complicate assessment, with pass@k metrics and verifier assistance substantially affecting reported performance.
• →Energy costs and inference efficiency are emerging constraints for reasoning-scale models, affecting practical deployment viability.
• →Future directions emphasize verified-discovery workflows that combine generation with formal verification for trustworthy AI-assisted mathematics.