Self-Consistent Decoding for More Factual Open Responses
This work addresses the problem of generating more factual open responses for users of large language models, representing an incremental improvement over existing decoding methods.
The authors tackled the problem of improving factuality in open response generation by extending self-consistency to decoding, using a 'Sample & Select' method that selects sentences based on token overlap. They showed a 30% relative improvement in factuality over existing decoders on NLI-based evaluations of CNN/DM and XSum subsets, while maintaining comparable ROUGE-1 F1 scores.
Self-consistency has emerged as a powerful method for improving the accuracy of short answers generated by large language models. As previously defined, it only concerns the accuracy of a final answer parsed from generated text. In this work, we extend the idea to open response generation, by integrating voting into the decoding method. Each output sentence is selected from among multiple samples, conditioning on the previous selections, based on a simple token overlap score. We compare this "Sample & Select" method to greedy decoding, beam search, nucleus sampling, and the recently introduced hallucination avoiding decoders of DoLA, P-CRR, and S-CRR. We show that Sample & Select improves factuality by a 30% relative margin against these decoders in NLI-based evaluation on the subsets of CNN/DM and XSum used in the FRANK benchmark, while maintaining comparable ROUGE-1 F1 scores against reference summaries. We collect human verifications of the generated summaries, confirming the factual superiority of our method.