Study finds LLMs can silently change a theorem’s scope when translating mathematics
An October 2 preprint tests whether AI preserves the assumptions hidden inside mathematical language.
When a mathematically fluent rewrite changes the claim
An October 2 preprint asks a less familiar question about AI mathematical reasoning: can a model preserve the scope of a statement when it rewrites it in the language of a neighbouring field? A result about vector spaces, for example, may need an explicit field assumption when recast as a statement about modules. Leaving that assumption unstated can turn a faithful-looking rewrite into a claim about a wider class of objects.
The researchers tested seven models across 90 paired statements, separating truth, content and scope rather than scoring one overall answer. When translating toward a more general formulation, the models widened the domain in 60.6% of rewrites; in the reverse direction, they narrowed it in 28.3%. The paper reports the pattern again on a held-out set and on statements written by mathematicians.
This is a study of a specific translation task, not a verdict that every AI-generated proof or explanation is unsound. But it gives users of mathematical AI a concrete review step: ask which assumptions were implicit in the original language and whether the rewrite has carried them across. The findings are a preprint and have not been independently replicated by MathsAI.