CGES: Confidence-Guided Early Stopping for Efficient and Accurate Self-Consistency
CGES:面向高效准确自一致性的置信引导早停方法
机构 * University of Massachusetts Amherst(马萨诸塞大学阿姆赫斯特分校)
专题命中 推理与问题求解 :large language model(abstract);language model(abstract);分类 cs.CL
AI总结 提出贝叶斯框架CGES,通过自适应停止采样减少自一致性推理调用次数,在5个推理基准上平均减少58%调用且精度损失仅0.4个百分点。
Comments Extended version. A preliminary version was accepted at the Efficient Reasoning Workshop @ NeurIPS 2025. Code: https://github.com/EhsanAghazadeh/cges