English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Large language models (LLMs) have demonstrated impressive capabilities instoring and recalling factual knowledge, but also in adapting to novelin-context information. Yet, the mechanisms underlying their in-contextgrounding remain unknown, especially in situations where in-context informationcontradicts factual knowledge embedded in the parameters. This is critical forretrieval-augmented generation methods, which enrich the context withup-to-date information, hoping that grounding can rectify the outdatedparametric knowledge. In this study, we introduce Fakepedia, a counterfactualdataset designed to evaluate grounding abilities when the parametric knowledgeclashes with the in-context information. We benchmark various LLMs withFakepedia and discover that GPT-4-turbo has a strong preference for itsparametric knowledge. Mistral-7B, on the contrary, is the model that mostrobustly chooses the grounded answer. Then, we conduct causal mediationanalysis on LLM components when answering Fakepedia queries. We demonstratethat inspection of the computational graph alone can predict LLM grounding with92.8% accuracy, especially because few MLPs in the Transformer can predictnon-grounded behavior. Our results, together with existing findings aboutfactual recall mechanisms, provide a coherent narrative of how grounding andfactual recall mechanisms interact within LLMs.

A Glitch in the Matrix? Locating and Detecting Language Model Grounding with Fakepedia