Tiejin Chen and 7 other authors submitted a preprint on 26 Oct 2026 titled "Detect Before You Disambiguate: Perturbation-Based Input Ambiguity Detection in LLMs."[1] On four question-answering datasets, the methods beat baselines that measure uncertainty in the model's answers or rewrite prompts with model paraphrases, the authors report.[1]
Perturbation-based LLM ambiguity detection beats baselines, authors report
On four datasets, the authors report, the methods beat baselines that measure uncertainty in the model's answers or rewrite prompts with model paraphrases.
1 primary · 0 independent · 0 commentary · 0 vendor