VecCISC: A Lightweight Framework for Efficient Weighted Majority Voting in LLMs

A new method called VecCISC reduces the computational cost of weighted majority voting in large language models (LLMs) by clustering semantically similar reasoning traces. Standard Self-Consistency selects the most common answer from multiple samples, while Confidence-Informed Self-Consistency (CISC) uses a critic LLM to assign confidence scores, improving accuracy but increasing overhead. VecCISC uses semantic similarity to filter redundant traces, lowering the number of critic calls. The approach is adaptive and lightweight, aiming to maintain performance gains while reducing expense. The paper is published on arXiv under ID 2605.08070.