LANTERN: Illuminating Hidden Mathematical Knowledge in Language Models
LANTERN uses model activations to surface OEIS relations, verifying 62 links including four claimed novel ones.
LANTERN uses a classifier on pretrained language-model activations to rank candidate mathematical relations, then filters, generates hypotheses, and checks them with executable and analytical verification. On the On-Line Encyclopedia of Integer Sequences it ranked 50 million pairs among 10,000 frequently referenced sequences and produced 62 verified relations that lacked an existing OEIS cross-reference. A content screen retained 13 relations, nine judged informative, including four the authors say are novel. The full pipeline finished in under eight hours.
- A classifier over pretrained activations ranks candidate integer-sequence relations.
- LANTERN scored 50 million pairs among 10,000 frequent OEIS sequences.
- It verified 62 cross-references; a screen kept 13, including four claimed novel.
- End-to-end training, ranking, filtering, and verification finished in under eight hours.
- Organizations
- OEIS
From the original article
1 min read at huggingface.co
Language models can now prove theorems, but people still decide which problems to pursue. We ask whether a model's internal representations can help identify promising mathematical connections.
By Pavel Tikhonov, Elena Tutubalina, Ivan Oseledets · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.