EE Seminar: LLM4BGP: Experimenting with LLMs for Networking Expert Systems
https://us04web.zoom.us/j/71362319075?pwd=B2kJq0yrqFrLCHlmmDemYqbxSRdhRz.1
Password: LLM4BGP
Electrical Engineering Systems ZOOM Seminar
Speaker: Dror Pezo
M.Sc. student under the supervision of Prof. Yuval Shavitt
Monday, 16th March 2026, at 15:00
LLM4BGP: Experimenting with LLMs for Networking Expert Systems
Abstract
Building a reliable "networking oracle", or networking "expert system" is difficult since its requires both knowledge of processes and data of various types, which is often dynamic or quasi-dynamic. The relevant available data sources on the Internet differ in structure, update cadence, and reliability, and frequently contain missing, stale, or conflicting information.
This thesis investigates large language model (LLM) agents as a basis for routing-oriented expert systems. We use an off-the-shelf (i.e., neither trained nor fine-tuned) LLM as a planning and reasoning component whose outputs are constrained to be grounded in evidence retrieved from authoritative sources.
Our results reveal a sharp split between retrieval and reasoning. For single- and multi-source fact retrieval, prompt-only approaches achieve low accuracy (0–45%), indicating that foundation models are unreliable for fast-changing Internet metadata. In contrast, tool-augmented agents reach 96–100% accuracy by retrieving and verifying up-to-date evidence.
We further deepened the study on Type-of-Relationship (ToR) inference and developed a multi-agent workflow that aggregates evidence from structured sources and performs targeted web retrieval of operator-facing documentation. Web augmentation improves end-to-end accuracy from 72.0% to 87.6%, highlighting the practical value of combining structured measurements with carefully vetted unstructured evidence.
Overall, the thesis provides empirical evidence and a concrete methodology for building routing-oriented expert systems with LLM agents. The results indicate that such systems can serve as a practical and transparent decision-support layer for routing analysis, enabling reproducible answers with explicit evidence trails while reducing dependence on continuous manual expert intervention.
-סמינר זה ייחשב כסמינר שמיעה לתלמידי תואר שני ושלישי-
This Seminar Is Considered A Hearing Seminar For Msc/Phd Students-
כדי לקבל קרדיט שמיעה יש לחתום שם מלא ומספר ת.ז. בצ'ט
