BioAlchemy: Extracting Reasoning Training Data from Biological Literature to Build Professional Scientific Reasoning Models

This article proposes the BioAlchemy workflow, which extracts verifiable scientific reasoning questions from biological research literature, constructs a professional dataset of 345,000 entries, trains the BioAlchemist-8B model via topic alignment and reinforcement learning, and achieves a 9.12% improvement on biological benchmark tests. This work addresses the lag in AI reasoning for biology and provides new insights for the field of scientific AI.

Biological data is abundant, but reasoning models perform worse on biological tasks compared to mathematics and programming fields. The core reason is that the theme distribution of existing reasoning datasets is seriously misaligned with modern biological research (focused on classic themes, lacking coverage of cutting-edge areas), leading to performance degradation when models handle practical problems. Additionally, extracting verifiable questions from biological literature faces challenges such as complexity, context dependence, and verification difficulties.

The BioAlchemy workflow includes steps like literature screening, question generation, answer extraction, verifiability checks, and diversity assurance. Its key innovation is explicit topic alignment (analyzing journal trends, identifying emerging areas, adjusting sampling weights). The constructed BioAlchemy-345K dataset features large scale (345,000 entries), diversity (covering multiple subfields), verifiability (with clear evidence), and topic alignment.

The BioAlchemist-8B model (8B parameters), trained using the BioAlchemy-345K dataset, employs reinforcement learning, focusing on reasoning chain generation, biological knowledge application, and cross-domain integration. Evaluations show that the model achieves a relative improvement of 9.12% on biological benchmark tests, with good cross-task generalization, especially notable improvements in topic-aligned tasks.

Core contributions of BioAlchemy: Revealing the dataset theme misalignment problem; developing a literature-to-training-data transformation workflow; constructing the 345K dataset; training a professional model with improved performance. This work emphasizes the importance of domain-specific data, verifiability, and topic alignment for scientific AI.

Application Prospects: Assisting biological research (literature review, hypothesis generation, experimental design), education (personalized learning), and interdisciplinary collaboration. Limitations: Some questions require expert verification, insufficient coverage of subfields, and limited complex reasoning capabilities. Future Directions: Expanding to other scientific fields, multimodal integration, real-time knowledge updates, and human-machine collaborative reasoning.