Reproduce “Old Ideas, Novel Problems: The Instability of LLM-Based Novelty Evaluation” 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
Noy Sternlicht, Simra Shahid, Peter Jansen, Daniel S. Weld, Pao Siangliulue, Tom Hope
The authors test how language model judges assess novelty in research ideas by comparing prompt formulations on the same pairs across human-only and human-plus-generated setups. They find that GPT-5.4’s pairwise strict accuracy on human-plus-generated pairs falls from 93.5% to 40.9% when reviewer framing is removed; the prompts disagree on 55.8% of pairs, showing prompt sensitivity, not verified novelty labels.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.