TL;DR
While AI has made strides in proving mathematical theorems, deciding what to prove remains a human-dominated art. This paper introduces the Theorem Evolution Graph (TEG), a structured framework that allows LLM agents to generate high-quality conjectures by explicitly mapping how existing theorems modify, strengthen, or extend one another.
What Problem Does This Paper Address?
In the world of automated mathematics, "verification" is the easy part—a formal proof can be checked mechanically. However, "conjecturing"—the act of proposing a new, meaningful mathematical statement—is notoriously difficult for AI.
The challenge isn't just generating a statement that might be true; it’s generating one that is worth pursuing. A conjecture must be novel, significant, and insightful. Until now, LLMs tasked with this have often produced "safe" but trivial variations of existing work. The authors address this by asking: Can we guide AI to find the "next logical step" in a research thread by encoding the history of how theorems evolve?
Methods
The researchers developed a system where an LLM agent (using Claude 3.5 Opus) constructs a Theorem Evolution Graph (TEG).
- Graph Representation:
- Nodes: Represent individual theorems, broken down into a canonical form: Setting ($\sigma$), Hypotheses ($H$), and Conclusion ($\kappa$).
- Edges: Describe the specific mathematical "move" made between theorems (e.g., "weakened the hypothesis" or "extended the result from finite groups to all groups").
- Agent Workflow:
- Identify: Starting from a "seed paper," the agent follows citations backward (to find foundations) and forward (to find extensions).
- Compare: The agent analyzes pairs of theorems to record exactly how one builds on the other.
- Generate: By looking at these "evolutionary chains," the agent identifies patterns—like a sequence of increasingly general settings—and proposes a new conjecture that continues that trajectory.
The study compared this TEG-based approach against an unconstrained baseline, where the agent could search and take notes however it chose without the graph structure.
Datasets
The authors evaluated their method on a test set of 30 seed papers:
- 12 Expert-Selected Papers: Chosen by a collaborating mathematician who authored them, ensuring high-quality ground truth for evaluation.
- 18 Math-Graph Papers: Selected from the Math-Graph dataset (Alexander et al., 2026), focusing on discrete mathematics and theoretical computer science.
Key Insights and Ideas
- Mathematics as Evolution: The core "aha moment" is treating mathematical progress not as a collection of facts, but as a directed graph of incremental changes. By forcing the LLM to explicitly label how a theorem changed a previous result, the model gains a "vector" for where to go next.
- The Ambition-Plausibility Trade-off: The study revealed a fascinating tension. TEG-generated conjectures were consistently rated as more ambitious and significant, while baseline conjectures were rated as more plausible. In other words, the graph encourages the AI to take bigger, more meaningful "leaps," even if those leaps are riskier.
- Structure Over Content: Ablation studies showed that even when the agent was given the exact same set of papers, it performed better when forced to organize them into a TEG. This suggests that the structured reasoning of the graph is more important than simply having access to the literature.
Key Results
The researchers used both LLM judges (GPT-4, Grok, Claude) and a blind expert mathematician to evaluate the conjectures.
- Preference: Judges preferred TEG-generated conjectures in 63% of non-tied cases.
- Expert Validation: The expert mathematician preferred the TEG conjecture for 8 out of 12 of their own papers (a 67% preference share).
- Qualitative Metrics: TEG conjectures significantly outperformed the baseline in Significance ($p < 0.001$), Ambition ($p < 0.001$), and Insight ($p = 0.002$).
- Case Study: In a test involving "Wythoff’s Nim" (a mathematical game), the baseline agent suggested a minor refinement. The TEG agent, however, identified a gap between two different papers and proposed a unified conjecture that applied to a much broader family of games.
Limitations and Future Work
- Truth vs. Value: The study measures if a conjecture is interesting, not if it is true. Proving these new conjectures remains a task for future work.
- Scale and Scope: The test set was limited to 30 papers, primarily in discrete math and CS. It remains to be seen if TEGs scale to more abstract fields like algebraic geometry.
- Hallucination Risks: While the TEG structure helps, LLMs can still misinterpret theorem statements or invent citations.
- Future Directions: The authors suggest integrating TEGs with Reinforcement Learning (RL). An agent could be rewarded for proposing conjectures that survive a "counter-example search" or a formal proof attempt, creating a closed-loop system for mathematical discovery.
Summary updated