How Open Must Language Models be to Enable Reliable Scientific Inference?
A new paper explores the implications of language model openness for scientific research, arguing that closed models generally hinder reliable inference. The authors propose that researchers must identify and justify their model choices, detailing steps taken to mitigate potential threats to inference. The paper suggests that current closed models are largely unsuitable for scientific use, with exceptions requiring careful consideration. AI
IMPACT Raises critical questions about the suitability of closed-source language models for scientific research and inference.