Genomic models learn regulatory code through in-context examples

Hao et al. prompted the Evo 2 genomic language model with 4-64 examples of functional DNA sequences to generate new regulatory candidates (Decoding Bio). The model identified specific attention components to recognize repeated patterns within these prompts, improving its design quality without additional training. The researchers then applied reinforcement learning to reward sequences predicted to have higher cell-type specificity, allowing the model to outperform specialized generative tools under limited data budgets.

This method relies on in-context learning, where the model adapts to a task by observing patterns in provided examples rather than updating its internal parameters. While the approach depends on the accuracy of existing biological prediction models, it demonstrates how AI can translate complex genomic structures into functional designs. This process mirrors the practice of documenting research findings as you work, which clarifies gaps in understanding by forcing a transition from raw data to a coherent, explainable output (Bri Does Things from The Curious Mind).