How BUPT NIRC’s Telco-Agent Won Global AI Telecom Challenge and Raised Autonomous Network Standards

Competition Background

During MWC Shanghai 2026, GSMA announced results of international telecom AI evaluations. The BUPT Network Intelligent Research Center (NIRC) team won the Open Telco AI Workshop & Hackathon global championship and placed second in the Telco Troubleshooting Agentic Challenge among 1,400 developers.

Event Overview

Open Telco AI Workshop & Hackathon : First GSMA event evaluating Telco Agents in real telecom scenarios, with 12 elite teams competing.

Telco Troubleshooting Agentic Challenge : Focused on end‑to‑end fault‑root‑cause identification and automated remediation, attracting 1,400 top developers.

Solution Highlights

The NIRC team introduced “Telco‑Agent: Hierarchical Cognitive Decision Architecture”. It addresses four fatal pain points of native LLMs in telecom operations: contextual drift, black‑box diagnosis, logic shortcuts, and hallucinated evidence.

Layered Task Decoupling

Recognizing that a single LLM cannot simultaneously perform root‑cause inference and decision making, the architecture separates diagnosis and optimization layers. The diagnosis layer converts over 90 high‑dimensional physical signals into 34 business‑semantic metrics and uses a Random Forest‑Decision Tree (RF‑DT) model for white‑box precise inference.

SOP Constraints and Evidence‑Driven Retrieval

To prevent premature convergence and logical shortcuts, expert SOPs are embedded into the LLM reasoning. A “Retrieval Gate” forces the model to query only minimal relevant data slices, cutting off the source of hallucinated evidence.

Ablation Results

Experiments show the combined “rules + data skills” architecture raises diagnosis accuracy from the raw 31.5 % to 100 %, achieves 0.92 decision‑layer accuracy, reduces latency by 44 %, and cuts token consumption by over 60 %.

Competition Experience

Facing strong global teams and a fully black‑box final stage, the NIRC team maintained top performance in preliminary rounds and secured the overall second place in the troubleshooting challenge.

Implications and Future Work

The results demonstrate the feasibility of deploying large models for end‑to‑end telecom fault remediation when tightly coupled with expert systems and rigorous engineering constraints. The team plans to continue collaborating with GSMA and industry partners to advance standards and translate research into production.