BiEvo: From autoregression to bidirectional genomic representations with MNTP and architectural symmetrization

Jan 1, 2026·
Feiying Chen
Feiying Chen
· 1 min read
BiEvo model overview.
Abstract
BiEvo is an ongoing project exploring how genomic sequence models can move beyond strictly autoregressive objectives toward bidirectional representations through masked-next-token prediction and architectural symmetrization.
Type
Publication
Ongoing manuscript
publications

Citation. Feiying Chen et al. BiEvo: From autoregression to bidirectional genomic representations with MNTP and architectural symmetrization. Ongoing manuscript.

This project studies bidirectional genomic representation learning, using objective and architecture changes to improve how sequence models encode biological context. It is listed as ongoing rather than under review.

Genomic Foundation Models Representation Learning AI for Biology
Feiying Chen
Authors
Feiying Chen
PhD Student in Biology
Feiying Chen is a PhD student in Biology at Tsinghua University. His research interests include perturbation biology, AI virtual cell modeling, single-cell transcriptomics, AI-driven drug discovery, and phenotype-based drug design.