Reproduce “Tracing Knowledge and Reasoning in LLM” 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

Yang Wang, Xiao Wang, Yiqi Liu, Chenghao Xiao, Eleni Vasilaki, Goran Nenadic, Chenghua Lin

The authors examine how language models’ internal processing relates to benchmark demands and whether multiple-choice answers depend on offered options. They analyze internal signals and perturb answer choices. On SciQ, accuracy drops more when the correct option is removed than when a wrong “none of the above” option is added, indicating option dependence in this test, not a universal cause.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.