This episode explores how Google researchers are tackling the issue of "hallucinations" in Large Language Models (LLMs) by connecting them to Data Commons, a vast repository of publicly available statistical data.
https://datacommons.org/

The researchers experiment with two techniques:
Retrieval Interleaved Generation (RIG), where the LLM is trained to generate natural language queries to fetch data from Data Commons and Retrieval Augmented Generation (RAG), where relevant data tables from Data Commons are added to the LLM's input prompt.

The paper details how these methods were implemented and evaluated using Google's Gemma and Gemini LLMs. Results indicate that both RIG and RAG show promise in improving the accuracy of LLM-generated statistical information, essentially giving LLMs a reality check. The authors also acknowledge limitations and outline future research directions, including expanding Data Commons' coverage and refining the training process for LLMs.

You can read the full paper here:
https://docs.datacommons.org/papers/DataGemma-FullPaper.pdf