Calibrated sequential stopping

Develop and evaluate a calibrated sequential stopping test based on the vote gap to adaptively terminate misi query processing while improving on the static (k_s,C) recall-cost frontier.

Background

The paper evaluates several adaptive stopping rules, including vote-set stability, tolerant stability, and verified-top-k patience, and finds that none improves on static parameter choices. It specifically identifies a calibrated sequential test based on the vote gap as an untested approach that could exploit the model’s signal without relying on the unsuccessful inexpensive stopping signals.

References

A calibrated sequential test on the vote gap --- the one variant the model does not forbid --- is untested and listed as future work.

— misi: a Metric Inverted Sample Index  (2608.27422 - Chavez, 27 Aug 2026) in Section 4, “What query-time processing does not add,” paragraph “Adaptive stopping”; Section 5, “Position”