Large Language Models Favor Biological Solutions Through Bio

A recent study published on arXiv examined the bias of large language models (LLMs) trained on extensive internet datasets, revealing a tendency towards favoring synthetic solutions over biological ones. By employing a Kelly criterion-inspired framework, the research analyzed ten LLMs with 50 curated prompts related to bioalignment across diverse domains including materials and energy. The findings indicated that most models displayed a significant bias against bioaligned outcomes, prompting further exploration into potential corrections through fine-tuning techniques.

The implications of this research are substantial for AI safety and application in bioengineering fields. Fine-tuning two specific models, Llama 3.2-3B-Instruct and Qwen2.5-3B-Instruct, with a focused corpus of biological articles led to a notable increase in their score for biological approaches, without detracting from their overall performance. This indicates that targeted fine-tuning could influence LLMs to adopt more bio-focused perspectives, potentially reshaping their deployment in industries reliant on biological solutions. The open release of benchmarks and model adaptations further signifies the study's commitment to advancing bio-oriented LLM architectures.