Perturbation-Based Ambiguity Detection for LLM Inputs
October 28, 2026
The proposed framework uses label-preserving perturbations to detect when an LLM input is ambiguous before the model attempts to answer. This approach outperforms existing output-side uncertainty methods by identifying upstream ambiguity using model-internal signals.
HOW THIS AFFECTS YOU
●
builderImplement these detectors to trigger clarification prompts when users provide ambiguous instructions.
●
researcherThis method provides a more robust way to quantify input uncertainty without relying on LLM-generated text.