Liang Wang

This paper re-analyzes a foundation model's checkpoints to show biological continued pretraining improves general knowledge, code generation, and reasoning without catastrophic forgetting, but subsequent fine-tuning reverses these gains.

The paper investigates whether structural transfer between natural language and biological foundation models is directional, finding that transfer from language to biology is stronger than the reverse.

This paper shows that general-purpose language models can detect protein homology without biological pre-training, achieving ROC-AUC 0.79 on a controlled benchmark.

OmniGene-4 is a multi-modal bio-language MoE model that decomposes task performance by router states, achieving strong benchmarks in homology and biological-knowledge tasks while being computationally efficient.

The paper introduces BioRAG-DRAG, a local-first multimodal retrieval system for biomedical agents, combining neural retrieval, BLAST verification, and graph-based evidence packaging.

The paper introduces Open-Rosalind, a tool-first biomedical agent system with a benchmark evaluating task accuracy, tool correctness, and trace completeness, showing tool-first execution improves performance.

This study uses large language models to rediscover the central dogma by aligning DNA and protein sequences, achieving 81% accuracy in a classification task.

The paper introduces a method to structure the human genome as a book by mapping DNA sequences to words, sentences, and paragraphs using transfer learning from English language models.