Reproduce “Black-Box Uncertainty Quantification for Large Language Models via Ensemble-of-Ensembles” 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

Wang Ma, Debarun Bhattacharjya, Junkyu Lee, Nhan H Pham, Harsha Kokel, Qiang Ji

The authors estimate language model uncertainty by repeatedly sampling answers to a prompt and to meaning-preserving rephrasings. Within-prompt embedding variation estimates aleatoric uncertainty, while variation across rephrased prompts estimates epistemic uncertainty. Controlled question-answering interventions found the scores respond differently to clarification and added evidence. On TriviaQA, AU achieved 0.789 AUROC at 25 generations, with overlapping confidence intervals.

Summary by alphaXiv.

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