To evaluate performance, the researchers tested Lean Conjecturer on 40 Mathlib seed files. From these inputs, the system generated 12,289 conjectures.
Lean Conjecturer: How AI Is Transforming Mathematical Discovery
Mathematics has long depended on human creativity and intuition. However, artificial intelligence is now entering this space. A new system called Lean Conjecturer shows how machines can help generate new mathematical ideas automatically.
Naoto Onda, Kazumi Kasaura, Yuta Oriike, Masaya Taniguchi, Akiyoshi Sannai, and Sho Sonoda conducted this research and published it under the title “LeanConjecturer: Automatic Generation of Mathematical Conjectures for Theorem Proving” in June 2025.
ENTECH STEM Magazine has included this research in its list of Top 10 Mathematics Discoveries of 2025.
This system connects large language models with formal logic tools. As a result, it produces original mathematical conjectures that fit within modern proof systems. By combining AI with formal verification, Lean Conjecturer opens new paths for mathematical research, education, and automated reasoning.
What Lean Conjecturer Is and Why It Matters
The Core Innovation Behind Lean Conjecturer
Lean Conjecturer is a pipeline designed for automatic conjecture generation. To achieve this, it uses large language models (LLMs) to produce new mathematical statements. Furthermore, these statements are formatted specifically for use in formal theorem proving. Technically, the system works in tandem with Lean 4, a proof assistant, and Mathlib, a large formal mathematics library. Ultimately, these tools work together to verify that the generated conjectures are both valid and meaningful.
Unlike simple text generation, Lean Conjecturer blends rule-based context extraction with AI-based generation. Because of this hybrid approach, the system produces conjectures that mathematicians can realistically study and prove.
The main motivation behind Lean Conjecturer is a known challenge in automated reasoning. Moreover, formal proof systems lack enough high-quality training data. Therefore, by generating new conjectures from existing mathematics, the system helps fill this gap.
How Lean Conjecturer Generates New Conjectures
At the center of Lean Conjecturer is the Mathlib library. First, the system reads existing Mathlib files. Then, it extracts relevant mathematical context from those files.
Next, an LLM generates new statements that resemble formal theorems. By doing so, it ensures these statements follow Lean 4 syntax, thereby making them immediately usable in proof environments. As a result, researchers can transition directly from AI-generated ideas to rigorous mathematical verification.
After generation, each conjecture goes through several validation steps:
- It must parse correctly in Lean 4.
- It must not already exist in Mathlib.
- It must be non-trivial and not easily provable.
Through this filtering process, Lean Conjecturer produces conjectures that genuinely expand formal mathematics.
Testing Results and System Performance
Large-Scale Conjecture Generation
To evaluate performance, the researchers tested Lean Conjecturer on 40 Mathlib seed files. From these inputs, the system generated 12,289 conjectures.
Out of these:
- 10,950 were valid formal statements
- 4,130 were new and novel
- 3,776 could not be proven automatically
On average, Lean Conjecturer created 103 new conjectures per file. In effect, this scale shows how AI can expand mathematical content far beyond manual human effort. Ultimately, these results clearly demonstrate that Conjecturer can generate large volumes of meaningful mathematical ideas while maintaining the rigorous standards required by the research community.
Improving Theorem Proving With Reinforcement Learning
Beyond generation, the team tested how these conjectures improve automated theorem proving; consequently, they applied Group Relative Policy Optimization (GRPO), a reinforcement learning method.
This training helped a theorem prover learn from the generated conjectures. Over time, the model improved its ability to solve difficult proof tasks.
As a result, the Lean Conjecturer not only produces ideas but also strengthens the systems that attempt to prove them. This feedback loop marks a major step forward in machine-assisted reasoning.
Real-World Impact of Lean Conjecture
Practical Uses of AI in Mathematics
Although Lean Conjecturer does not replace human insight, it provides valuable support. Researchers can use the system to:
- Generate new training data for theorem provers
- Explore underdeveloped mathematical areas
- Reduce manual conjecture writing
- Accelerate formal mathematical discovery
In this way, Lean Conjecturer helps mathematicians focus on deep reasoning rather than repetitive tasks.
Applications Beyond Pure Mathematics
The impact of Lean Conjecturer extends beyond theory. For example:
- Software verification benefits from stronger formal proofs.
- Education gains new examples for teaching logic and proofs.
- AI reasoning systems improve through better logical training.
By automating part of discovery, researchers can redirect effort toward creative and conceptual challenges.
Adoption Timeline and Future Outlook
Currently, Lean Conjecturer remains a research system. Still, the results show that large-scale conjecture generation is already feasible.
With further integration into proof environments and better interfaces, systems like Conjecturer could see broader academic use within the next few years. Over time, such tools may become standard components of mathematical workflows.
Research Areas and Career Paths for Students
Fields to Study
Basically, this work highlights several promising study areas:
- Artificial Intelligence and Machine Learning
- Formal Logic and Theorem Proving
- Computer Science and Software Verification
- Pure and Applied Mathematics
Each field benefits from the ideas behind Conjecturer.
Career Opportunities
Students can pursue roles such as:
- AI Research Scientist
- Software Engineer in verification systems
- Mathematics researcher using computational tools
- Data scientist working on reasoning systems
Skills in logic, computation, and creativity are valuable across these careers.
A New Era of Machine-Assisted Mathematics
Broadly speaking, Lean Conjecturer proves that AI can assist with deep intellectual tasks once reserved for humans. By effectively generating thousands of formal mathematical ideas, it not only expands the scope of discovery but also redefines the relationship between human intuition and machine logic. Consequently, we are entering an era where AI serves as a tireless partner in the pursuit of abstract truth.
By integrating Lean 4, Mathlib, and modern AI, the system not only bridges machine generation with formal proof but also enhances the overall process. Consequently, in the coming years, Lean Conjecturer may significantly reshape how mathematics is researched, taught, and applied across disciplines.
Additionally, to stay updated with the latest developments in STEM research, visit ENTECH Online. Basically, this is our digital magazine for science, technology, engineering, and mathematics. Further, at ENTECH Online, you’ll find a wealth of information.
Reference:
- Onda, N., Kasaura, K., Oriike, Y., Taniguchi, M., Sannai, A., & Sonoda, S. (2025). LeanConjecturer: Automatic generation of mathematical conjectures for theorem proving. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2506.22005
- Author
- Latest Posts
I’m a web developer and science writer with a Master’s degree in Computer Applications (MCA) from Pune University.
I work in tech, building websites and staying updated with the latest developments. I also love writing about physics, chemistry, technology, robotics, mathematics, and breakthrough innovations. My passion is breaking down complex scientific topics into simple, easy-to-understand stories.
My goal is to help everyone stay informed about the exciting changes happening in science and technology. I believe understanding science shouldn’t be complicated—it should be accessible to all.
