The conjecture that natural-language-trained large language models cannot reliably reason at professional mathematics

Large language models (LLMs) are systems trained to generate text from natural-language data. Natural-language reasoning conjecture. Large language models (LLMs) trained on natural language alone will not reason reliably at the level of professional mathematics. This motivates training AI systems on formal mathematical data and integrating them with interactive theorem provers; the source gives a counterpoint that some groups are actively working to disprove the claim, so its resolution remains open.