Reproduce “LLM Abstention Can Be a Prompt Artifact, in Addition to Genuine Uncertainty” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zipeng Ling, Shuliang Liu, Yuehao Tang, Junqi Yang, Shenghong Fu, Seonil Son, Chen Huang, Kejia Huang and 3 more
The authors test whether adding an Unknown option makes language models abstain from answerable questions. They compare True-or-False tasks with four-option multiple-choice questions across three models and two datasets. Adding Unknown lowers True-or-False accuracy by an average of 19.8 percentage points, versus 1.2 points for multiple-choice questions, so the effect depends on question format and the tested models and tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.