CGES: Confidence-Guided Early Stopping for Efficient and Accurate Self-Consistency
Researchers have developed a new Bayesian framework called Confidence-Guided Early Stopping (CGES) to improve the efficiency of large language model (LLM) querying. CGES adaptively halts sampling once a single answer gains sufficient confidence, unlike traditional self-consistency methods that require a fixed number of calls. This approach significantly reduces the number of LLM calls needed, cutting them by an average of 58% across five reasoning benchmarks, while maintaining accuracy comparable to the standard self-consistency strategy. AI
IMPACT Reduces computational cost for LLM inference, potentially enabling wider deployment of complex reasoning tasks.