UniEval

UniEval reframes text-generation evaluation as a Boolean question-answering task: a single T5-based model is asked one yes/no question per quality dimension (e.g. "Is this a coherent summary?"), and the probability it assigns to "Yes" becomes that dimension's score [@zhongUnifiedMultiDimensionalEvaluator2022]. Because each dimension is just a natural-language question, one unified model covers coherence, consistency, fluency and relevance, and it can extend to new dimensions by adding a question. This one-question-per-dimension design is the direct precursor to Ask, Don’t Judge: Binary Questions for Interpretable LLM Evaluation and Self-Improvement's BinEval, which keeps the binary-question idea but decomposes each dimension into many questions rather than one.