BARS method for budget-aware majority identification in LLMs
November 11, 2026
BARS is a sequential sampling rule that manages LLM self-consistency under a fixed budget. It certifies the top-two gap to decide between early stopping, a budgeted majority vote, or a refusal when confidence cannot be established within the error rate.
HOW THIS AFFECTS YOU
●
builderYou can use this to decide when to stop sampling to save costs or when to switch to a larger model.
●
researcherThis provides proven bounds on joint answering error and refusal rates for majority voting.