2 ms·
Challenges with unsupervised LLM knowledge discovery
- westurner 3y agoDo LLMs converge upon the same answer when asked the same question multiple times? If they do not converge, are any of the replies true? Or do they have high attention values?
- westurner 3y ago"Challenges with unsupervised LLM knowledge discovery" (2023) https://arxiv.org/abs/2312.10029 https://arxiv.org/abs/2312.10029 : > Abstract: We show that existing unsupervised methods on large language model (LLM) activations do not discover knowledge -- instead they seem to discover whatever feature of the activations is most prominent.
- westurner 3y ago- NewsArticle about / that repeats a ScholarlyArtice: "A New Research from Google DeepMind Challenges the Effectiveness of Unsupervised Machine Learning Methods in Knowledge Elicitation from Large Language Models" (2023) https://www.marktechpost.com/2023/12/20/a-new-research-from-google-deepmind-challenges-the-effectiveness-of-unsupervised-machine-learning-methods-in-knowledge-elicitation-from-large-language-models/ https://www.marktechpost.com/2023/12/20/a-new-research-from-...