CERET is a framework for refining LLM predictions by considering semantic stability, entailment and inter-sample uncertainty measures. This approach does not require additional training, or iterative inference of LLM(s).
Experimental results show that CERET significantly outperforms Self-consistency and Self-rarank baselines for abstractive summarization and question answering. Compared to various LLM self-improvement / self-reflection methods, our approach has lower latency and is more cost-effective.
