Model Editing for Generative Information Retrieval

AI Technology: Natural Language Processing

This project addresses the growing need for accurate information retrieval in light of challenges posed by large language models (LLMs) like ChatGPT, which can produce misleading content through a phenomenon known as hallucination. 

Traditional Information Retrieval (IR) systems, which are essential for fact verification in LLM-based systems, face trade-offs between storage requirements and retrieval speed. 

Generative Information Retrieval (GenIR) models improve efficiency by embedding the document corpus into a single Transformer model, allowing for fast and effective retrieval, but they struggle to adapt to changes in the corpus without time-consuming retraining.

To overcome these limitations, the project aims to enable dynamic editing of GenIR models, allowing for the addition, modification, or removal of documents post-training. 

By leveraging mechanistic interpretability and model editing techniques, the proposed project seeks to create editable GenIR models suitable for real-world applications.

Key objectives include understanding how document information is stored in GenIR models and developing reliable, non-destructive editing methods. 

The expected outcomes are a mechanistic understanding of document storage, scalable editing techniques, and benchmarks demonstrating the viability of editable GenIR systems, ultimately benefiting search and fact verification systems while improving the understanding of generative models.