Researchers have introduced LANTERN, a pipeline that asks whether the internal representations of a language model can help point mathematicians toward promising connections. The work, posted as a Hugging Face paper, focuses not on theorem proving itself but on the earlier, more human step of deciding which problems and relationships are worth pursuing.
The method trains a classifier over the activations of a pretrained model to rank candidate relations between objects, then applies staged filtering, hypothesis generation, executable verification, and analytical checking. According to the paper, this design is intended to be fast and cost-efficient rather than relying on heavy generation alone.
Applied to the On-Line Encyclopedia of Integer Sequences (OEIS), LANTERN ranked 50 million pairs drawn from 10,000 frequently referenced sequences. It produced 62 verified relations between pairs that had no existing OEIS cross-reference.
A subsequent content screen retained 13 relations deemed worth presenting. Of those, nine were judged informative or insightful, including four described as entirely novel to the best of the authors' knowledge — none appearing in OEIS or in their targeted literature search.
The authors report that the full end-to-end process — classifier training, candidate ranking, filtering, and verification — took under eight hours, suggesting that activation-based ranking could serve as a practical first pass for mathematical discovery.