LDARNet: DNA Adaptive Representation Network with Learnable Tokenization for Genomic Modeling

In three linesLDARNet, a 120M-parameter genomic foundation model, introduces learnable adaptive tokenization without supervision to replace fixed schemes (k-mers, BPE). Combining BiMamba-2 with local attention and bidirectional routing, it achieves state-of-the-art on 5 histone modification tasks and outperforms models 20× larger at matched compute.

Summary generated by Claude — human-verified