Reproduce “Specification-Aware Predictive Uncertainty Decomposition” 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
Roman Vashurin, Mikhail Seleznyov, Assel Yermekova, Fedor Chernogorskii, Danila Zykov, Elena Tutubalina, Alexander Panchenko, Preslav Nakov and 2 more
SEA separates predictive risk into specification, residual aleatoric, and epistemic uncertainty, linking them to clarification, irreducible outcome risk, and model limitations. It estimates specification uncertainty from possible user intents and decomposes risk under proper scoring rules. Across tested text, image, and audio question-answering settings, it was the strongest signal for detecting underspecified inputs, not evidence that clarification improves answers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.