Semantic Routing Calibration to Mitigate LLM Over-Refusal
September 23, 2026
Semantic Routing Calibration (SRC) is a training-free inference framework that suppresses hypersensitive safety heads in transformers. It prevents over-refusal by addressing internal routing conflicts where benign prompts are incorrectly bound to refusal semantics.
HOW THIS AFFECTS YOU
●
builderYou can reduce false-positive safety refusals in your applications without retraining your models.
●
policyThis provides a mechanistic way to tune safety behavior during inference.