Harnessing Black-Box Control to Boost Commonsense in Language Models' Generation

Explore a resource-efficient framework for enhancing commonsense in large language models during a 55-minute talk presented by Yufei Tian from UCLA at the USC Information Sciences Institute. Discover the BOOST method, which steers frozen Pre-Trained Language Models towards more reasonable outputs without expensive fine-tuning. Learn about the creation of an interpretable, reference-free evaluator that assigns commonsensical scores to sentences based on a dynamic knowledge base. Examine how this evaluator guides the NADO controllable generation method to train an auxiliary head, improving output quality. Review test results on various language models, including GPT-2, Flan-T5, and Alpaca-based models, and compare BOOST-generated content with ChatGPT outputs through human evaluation. Gain insights into creative and controllable text generation, machine reasoning, and evaluation metrics for open-ended NLG tasks from Yufei Tian, a CS PhD student at UCLA supported by the UCLA-Amazon fellowship program.