Arc/Stanford: Evo 2 can design bacteriophage genomes from a handful of examples in the prompt

Gary Brixi and Brian Hie's teams report that the genome language model Evo 2 can learn from a few examples placed in its prompt (in-context learning, no retraining) and use them to generate proteins, de novo antibodies, CRISPR-Cas9 loci, mitochondrial genomes and entire bacteriophage genomes. Feeding the model's own synthetic sequences back into the prompt (a 'chain-of-thought') helps reach harder design goals. The work is the bioRxiv preprint "Efficient biological design via genomic in-context learning", posted 29 October 2026 and not yet peer reviewed.