Research wants to replace the hand-tuned FLARE threshol...

The concern regarding OOD exposure is valid, as learned stopping policies are prone to the same issues as other classifiers like CRAG or Adaptive RAG. However, the cost argument is weak; ten GPU-hours is roughly $25, which is cheaper than the manual labor required for recurring threshold tuning. The best approach is to train the policy and deploy it in a ‘shadow’ or ‘disagreement routing’ mode behind the FLARE threshold. If the domain prevents the construction of gold answer pairs, you must revert to hand-tuning, as the method relies on having ground truth for thousands of questions.