Schedule

This is a half-day (3-hour) lecture-style tutorial, including scheduled breaks. It will be conducted on-site, with at least two presenters attending the conference in person.

Introduction and Background (30 min)

• Tutorial Overview

• Foundations of Retrieval-Augmented Generation (RAG)

• Emerging RAG Paradigms

• Limitations of Standard RAG

Topic 1 — Dynamic RAG (50 min)

• Motivation: Why is Dynamic RAG necessary?

• When to Retrieve: When should the retrieval module be triggered during the generation process?

• What to Retrieve: How can we formulate queries that reflect the model’s real-time information needs?

• Open Challenges and Future Directions

Q&A (10 min)

Break (30 min)

Topic 2 — Parametric RAG (40 min)

• Background: Overview of existing works on how LLMs encode world knowledge in their parameters.

• Motivation: Scenarios where in-context knowledge injection faces limitations.

• Parametric Knowledge Representation: Methods for converting textual knowledge into plug-in parameters.

• Existing Works on Parametric RAG: Representative approaches and their design choices.

• Open Challenges and Future Directions

Summary and Future Directions (10 min)

A summary of the tutorial’s key insights, open problems, and future research opportunities in retrieval-augmented generation.

7. Final Q&A (10 min)