Skip to main content
Back to timeline
arXivSource publication:

Theorems no longer stand alone: a graph of how results build on each other yields conjectures judges prefer

Synopsis

Mathematical agents have mostly worked on proving existing conjectures, while deciding what is worth proving lacks an operational criterion. The Theorem Evolution Graph records how theorems build on one another: nodes are theorems, and edges capture changes in setting, hypotheses and conclusions. Agents search citations backward and forward from a seed paper, compare theorems, and propose conjectures from those differences. Across 30 seed papers, three model judges preferred graph-generated conjectures in 63% of non-tied judgments, and an expert preferred them on 8 of 12 papers; baseline conjectures were favored for plausibility.

Source-provided article image: Guiding AI Agents to Generate Conjectures by Encoding How Theorems Build on One Another
arXiv

Interpretation

Mathematical agents are now guided toward choosing what to prove, not only proving existing statements: they chain theorems around a seed paper into a directed graph along citations and propose new conjectures from the differences between those theorems. Earlier conjecture generation mostly treated conjectures as training data for theorem provers or produced formal conjectures from existing libraries; judging whether a conjecture is novel, significant or likely to deepen understanding had little structural basis. On 30 seed papers, three model judges produced 90 judge-paper judgments: 50 chose the Theorem Evolution Graph conjecture, 29 chose the unconstrained agent's, and 11 were ties; among non-tied judgments the graph conjectures took 63% (95% CI 52%–73%).

The agent first decides which theorems are related, then compares their settings, hypotheses and conclusions one by one, writing onto each edge what new idea the step introduced. An edge is neither a citation nor a proof dependency but a change in the theorem statement itself: a hypothesis relaxed, a conclusion strengthened, or a result extended to a new setting. Construction starts from the seed paper's main theorems, searches backward and forward through Semantic Scholar citation endpoints, and uses read_paper to read full texts and extract theorems; with the available papers held fixed, the agent building and using the graph drew 58% of non-tied judgments (95% CI 43%–72%), and 64% (50%–77%) against the original agent, though neither comparison reached statistical significance at this sample size (p = 0.36 and 0.07).

Perspective

The demonstrated scope is 30 seed papers with a single generation model, Claude Opus 5.5, evaluated mainly through model-judge preferences and one expert's blinded review of 12 of those papers; every generated conjecture remains unproved and is a candidate for expert review.

The authors note that the seed papers are few and selected and that only one generation model was used, so findings may not generalize across mathematical fields or models; theorem extraction and edge annotations were not validated by experts at scale, and errors in either can propagate into conjectures; judges may favor ambitious-sounding writing, and agreement between the expert and the model panel is moderate.