Researchers find Mythos activations behave like genomic language models, hinting LLMs learn DNA motifs

jatin_n0 · x · 2026-09-24

Joseph Kellogg observed that Mythos activations contain directions behaving similarly to genomic language models; jatinn0 hypothesizes that sufficiently large LLMs with sequences in pretraining may pick up coupling signals and motif-like features that pLMs/gLMs capture.
Stanford's Anshul Kundaje added context: he tried Claude/GPT for biologically annotating DNA motifs (what they do / bind to) and both were pretty bad out of the box (Opus 5 and GPT6 tested; Mythos inaccessible). He frames this as a multi-modal reasoning task where a trained LLM could unlock a massive bottleneck in genomic sequence annotation, a currently painstaking manual labeling effort.
The finding may also explain why some Claude models recognized ART arrays while others didn't.

More from Models