Large Language Models

This article discusses the transformative potential of using large language models as uncertainty-calibrated optimizers for scientific experimentation. It highlights how these models can enhance discovery processes in chemistry and materials science by effectively managing exploration and exploitation.

While AlphaFold revolutionized biology through massive datasets, most scientific fields lack the infrastructure to replicate such data-heavy models. The future of AI in science lies in autonomous agents capable of reasoning and utilizing digital tools to mimic human scientific inquiry.

The paper presents AISPA, a structured auditing taxonomy to identify harmful or manipulative instructions within hidden AI system prompts. Through an audit of 88 commercial products, the researchers found that despite growing trends in user protection, many systems still contain instructions that work against user interests.

Researchers have developed AISPA, a structured framework to audit the hidden system prompts that govern LLM behaviors. An analysis of 88 commercial products found that while developers are increasing protective instructions, many prompts still contain deceptive or manipulative directives.