A new arXiv preprint asks a pointed question: do genomic foundation models provide predictive information beyond simple sequence composition? The authors evaluate frozen representations from five DNA foundation models against normalized k-mer frequencies across 57 genomic classification tasks.

The design is straightforward: if k-mer counts—a basic measure of sequence content—match or beat the learned representations, then the models' added complexity may not be buying much predictive power. By testing five different models on a common set of tasks, the study aims for a systematic comparison.

However, the abstract as released stops before presenting results. It lays out the question and methodology but does not state whether the foundation models actually outperformed the k-mer baselines. Readers will need the full paper to learn the answer.