When Genomic Masking Priors Fail to Transfer: Strong Variant Prediction, Weak Functional Generation

arXiv published the paper “When Genome Masking Priorities Fail: Strong Variation Prediction, Weak Function Generation” on September 7, 2026, indicating that the entropy-guided placement assumption in gene composition modeling only holds partially for bidirectional discrete diffusion models. The supervised-fine-tuned GenDA model with 202M parameters achieved an AUROC of 0.774 for ClinVar SNV, outperforming self-regressive models of the same size with an AUROC of 0.103, but matching random span variation only at 0.777, proving that entropy guidance did not improve clinical variation prediction. The model performed poorly in zero-sample function filling tests, failing to consistently surpass the control group consisting only of 3-mer sequences at promoters, enhancers, and boundaries, and this failure existed at 50–500 bp gaps and worsened with increasing length. The analysis showed that entropy measures local complexity rather than functional importance; the 1-mer segmentation limit restricted physical context, and the training span上限 was 300 bp.