LANTERN: Illuminating Hidden Mathematical Knowledge in Language Models

The authors develop LANTERN, a fast and cost-efficient pipeline that ranks candidate mathematical relations using a classifier over pretrained model activations, followed by staged filtering, hypothesis generation, executable verification, and analytical checking. Applied to the On-Line Encyclopedia of Integer Sequences (OEIS), LANTERN ranked 50 million pairs among 10,000 frequently referenced sequences and produced 62 verified relations between pairs without existing cross-references. A content screen retained 13 insightful relations, including 4 entirely novel ones. The entire end-to-end process executed in under 8 hours.

LANTERN ranked 50 million pairs among 10,000 frequently referenced sequences from the On-Line Encyclopedia of Integer Sequences (OEIS).

The pipeline produced 62 verified relations between pairs without an existing OEIS cross-reference.

The entire end-to-end process took under 8 hours.