Selective intent routing gives an AI assistant a way to act on predictions it is sure about while deferring requests that are uncertain. The difficulty is knowing when a prediction is actually reliable. Standard confidence scores, the paper argues, primarily reflect the base model's representation, which may not fully capture the risk of acting on a wrong intent.
The proposed approach, signed lexical confidence, adds a lexical signal to the confidence estimate. This extra information is meant to make routing decisions risk-calibrated, so the assistant defers precisely in cases where acting would be costly. The abstract is truncated, so specific experimental results are not available in the source, but the motivation is clear: better confidence signals should lead to safer deferral behaviour.