Evo 2's Emergent In-Context Learning Draws Praise and Skepticism

Episode 1 · Evo 2 exhibits emergent in-context learning, enabling few-shot genomic design (2026-10-29, 14 posts)

Brian Hie's team (Arc Institute/Asimov) released a preprint reporting an unexpected emergent capability in the genomic language model Evo 2: generative in-context learning (ICL). Given just a few sequences sharing a common function in the prompt, the model can generate new sequences with similar functionality—no fine-tuning required—dramatically lowering the barrier to genome design.

Confirmed

• Core finding: feeding Evo 2 a small number of example sequences in the prompt enables it to generate new sequences with similar function, a form of LLM-like generative ICL; per @garykbrixi, after concatenating a few related sequences into the prompt, the model appears to learn an "algorithm" for recombining and mutating them on its own, working across proteins, RNA, and whole genomes (m1, m3, m5).
• Phage design: the Zenodo study "Efficient biological design via genomic in-context learning" shows that while phage design with Evo 2 previously required fine-tuning on thousands of genomes, in-context learning with just 4 examples now achieves the goal across multiple tasks, surpassing fine-tuned models without any fine-tuning (m4, m7).
• Enhancer design: according to Brian Hie (m9), with only dozens of example prompts, Evo 2 designed mouse in vivo-specific enhancers without additional training, achieving few-shot design of mammalian regulatory elements; the work comes with a paper, code, and a web app (m2).
• Antibodies and phages: LeoTZ03 relayed that few-shot examples can also steer the model to generate antibody sequences and phage genomes (m6).
• Genomic chain-of-thought (CoT): feeding screened designs back into the prompt builds a recursive "genomic chain-of-thought" that helps designs reach regions of sequence space inaccessible to either natural prompts or fine-tuning; the proportion of novel genomes rose from 0.25% to 2.9%, and the predicted capsid protein still folded like ΦX174 (m8).

Controversy and Skepticism

• Stanford professor Anshul Kundaje (m10) questioned whether designing short regulatory elements for reporter gene constructs really requires a generative model like Evo 2, while saying he's open to being convinced; he argued a more interesting direction would be pairing long-context Evo 2 generators with long-context supervised models to design complete genomic regions.

Why It Matters

• ICL is an emergent phenomenon not anticipated during training, bringing the LLM-style few-shot learning paradigm into genome design and sharply lowering the barrier to use (from fine-tuning on thousands of genomes down to single-digit examples).
• Genomic CoT demonstrates a recursive path to expanding reachable sequence space, potentially opening a new paradigm for genome modeling and reasoning—a focal point of public discussion among scholars including Kundaje (m2, m10).

• Evo 2 Shows Emergent In-Context Learning: Few-Shot Prompts Steer Design of Enhancers, Antibodies, Phage Genomes — BrianHie · 2026-10-29
• Evo 2 genome model designs cell-type-specific enhancers from just tens of examples — BrianHie · 2026-10-29
• Designing phages without fine-tuning: Evo 2 does it with four in-context examples — garykbrixi · 2026-10-29
• Evo 2 appears to learn recombination algorithms just by concatenating sequences in the prompt — garykbrixi · 2026-10-29
• Prompting Evo 2 with 14 examples: 15% of generations pass lanmodulin filters — garykbrixi · 2026-10-29
• Four natural loci in the prompt let Evo 2 generate full Cas9 systems at 10× less compute — garykbrixi · 2026-10-29
• Evo 2 generates complete Cas9 loci from 4 natural loci in prompt at ~10x less compute — garykbrixi · 2026-10-29
• Evo 2 genome design: 4-8 natural genomes in prompt lifts pass rate from 13% to 52% — garykbrixi · 2026-10-29
• Genomic chain-of-thought: novel genome share rises from 0.25% to 2.9% with Evo 2 — garykbrixi · 2026-10-29
• Evo 2 designs phage genomes via in-context learning, beating fine-tuned models — garykbrixi · 2026-10-29
• Genome Model Evo 2 Learns In-Context, Generating Antibodies and Phages — LeoTZ03 · 2026-10-30
• Evo 2 shows emergent generative in-context learning for few-shot regulatory sequence design — anshulkundaje · 2026-10-30
• Genomic in-context learning and CoT touted as a new frontier for generative biology — anshulkundaje · 2026-10-30
• Researcher questions whether short regulatory element design needs an Evo 2 generator — anshulkundaje · 2026-10-30

Episode 2 · Stanford Professor Questions EVO2: Simpler Models Suffice for Enhancer Design (2026-10-30, 2 posts)

Stanford professor Anshul Kundaje argues that simpler models can achieve enhancer and promoter design as well as or better than long-context DNA language models like EVO2.

• Stanford's Kundaje questions DNA LMs: simpler methods suffice for promoter design — anshulkundaje · 2026-10-30
• Stanford's Kundaje challenges EVO2: simpler models already handle enhancer design — anshulkundaje · 2026-10-30