The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement

Yi Duan^*, Ying Liu^*, Zirui Tang^*, …, Xuanhe Zhou, Fan Wu

Technical Report 2026
A roadmap toward genuine recursive self-improvement, organizing AI systems by how much responsibility they assume within the improvement loop—from executing predefined improvements to recursively improving the improvement process itself. Code

Biological Reasoning-Informed Regression for Interpretable Regulatory DNA Activity Prediction

Yi Duan^*, Zhao Yang^*, Jiwei Zhu, Ying Ba, Chuan Cao, Bing Su

KDD 2026
We design an interface R3LM that enables LLMs to directly understand DNA sequences, introduce CRE-ReasonBench, and develop a biological reasoning-informed regression framework for interpretable regulatory DNA activity prediction. Code

Extending Sequence Length is Not All You Need: Effective Integration of Multimodal Signals for Gene Expression Prediction

Zhao Yang^*, Yi Duan^*, Jiwei Zhu, Ying Ba, Chuan Cao, Bing Su

ICLR 2026, (Oral, 1.15% of submitted papers)

Through systematic experiments, we find that current gene expression predictors have limited long-sequence modeling ability. Prism integrates multimodal epigenomic signals with DNA sequences to achieve state-of-the-art performance in gene expression prediction. Code

BioMatrix: Towards a Comprehensive Biological Foundation Model Spanning the Modality Matrix of Sequences, Structures, and Language

Qizhi Pei^*, Zhimeng Zhou^*, Yi Duan^*, …, Lijun Wu

BioMatrix Technical Report, 2026

BioMatrix unifies molecules, proteins, sequences, structures, and language in a shared discrete token space. As a core co-first author, I completed 39 of 80 biological tasks, built the 1D molecule evaluation pipeline, and supported key results and representation analyses. Code

R³: End-to-End Reasoning-based Planning for Multi-step Retrosynthesis via Reinforcement Learning

Yifei Wang^*, Qizhi Pei^*, …, Yi Duan, …, Lijun Wu, Weiying Ma, Hao Zhou

ACL 2026, Main Conference

An end-to-end retrosynthesis planning framework that uses generative reasoning and reinforcement learning to replace conventional search-heavy multi-step synthesis planning.