Adaptive Self-Consistency Prompting for Cost-Efficient Reasoning in Large Language Models

This work examines Adaptive Self-Consistency, an
efficiency-focused prompting strategy for large language models,
evaluated on the GSM8K grade-school math benchmark. Rather
than drawing a fixed batch of reasoning traces up front, the
method samples reasoning paths one at a time and halts as
soon as the model outputs reach sufficient agreement, acting as
a decision boundary rather than committing to a fixed sample
count. The approach is compared against a plain single-pass
Chain-of-Thought baseline and against fixed Self-Consistency
with k = 10. A train/test split is used for threshold tuning and final
evaluation respectively, and both accuracy and the mean number
of model calls per question are reported as a cost proxy.