Genomic chain-of-thought: novel genome share rises from 0.25% to 2.9% with Evo 2
garykbrixi · x · 2026-10-29
Genome language models never see reasoning traces in training, but feeding filtered designs back into the prompt creates a genomic chain-of-thought. This recursive approach reaches regions of sequence space that neither natural prompts nor fine-tuning reach: the share of novel genomes rises from 0.25% to 2.9%, and predicted capsids still fold like ΦX174's.
Earlier in the thread: 4–8 natural ΦX174-like genomes in the prompt beat prior fine-tuning on thousands — 52% of designs pass correctness filters vs 13%, and a library of 64 novel genomes takes 2.1 GPU-hours instead of 200; just changing the prompt extends this to RNA phage MS2 and filamentous phage M13.
More from Research
- CVPR 2027 caps submissions at 25 per author and mandates LLM usage statements — CSProfKGD · 2026-10-30
- NUS's Xavier Bresson releases Lecture 10: GNNs Meet Large Language Models — xbresson · 2026-10-30
- One-neuron owl in Qwen3-8B fires on the chat template, not the persona — maxsloef · 2026-10-30
- WSDM 2027 to host LongSeek, first workshop on long-horizon information seeking — mrdrozdov · 2026-10-30
- Lean 4 formalization of an asymmetric metric-tensor physics paper hits GitHub — SubtleGradient · 2026-10-30
- Is sycophantic AI inflating research publications? Two mechanisms worth questioning — KLdivergence · 2026-10-30