Research
BARS: Budget-Aware Sequential Sampling with Refusal for LLM Majority Identification
Researchers propose BARS, a sequential sampling rule for majority identification under an unknown distribution with a fixed budget and target error rate. BARS certifies the top-two gap and returns either an early stop, a budgeted majority vote, or a refusal when feasibility cannot be established. In experiments, its practical mode draws 22.3% fewer samples than fixed-8 sampling on GSM8K while declining 18.8% of questions. On MATH, it reaches 82.5% answer accuracy while refusing half of difficult questions the model cannot solve.
large language modelsself-consistencyinference cost