TeleLogsAgent is a benchmark and evaluation framework designed to measure the ability of Large Language Model (LLM) agents to perform structured tool-use in the telecommunications domain.

It simulates the workflow of a 5G network engineer diagnosing performance degradation during drive testing, requiring agents to:

inspect configuration data,
analyze time-series KPIs,
reason across multiple tools,
identify the most plausible root cause.

Overview

The benchmark consists of two main components:

FastAPI Server (fastapi_server.py)
Exposes realistic analytical tools (HTTP endpoints) to access 5G drive-test scenarios.
Agents interact with this environment using OpenAI-style function calls.
LLM Evaluation Agent (benchmark.py)
Connects to either the FastAPI server and evaluates LLMs on their ability to plan, call tools, and reason over multiple steps.

In addition, we conveniently provide a FastMCP Server (fastmcp_server.py) as an alternative implementation of the FastAPI server using FastMCP. This version is especially convenient for MCP-native LLM agents.

Project Structure

TeleLogsAgent/
├── fastapi_server.py      # FastAPI benchmark server (HTTP tools)
├── fastmcp_server.py      # FastMCP benchmark server (MCP tools)
├── benchmark.py           # LLM evaluation / benchmarking script
├── TS1/test.json          # Scenario 1: root cause identification based on high-level network configuration and user-plane data.
├── TS2/test.json          # Scenario 2: root cause identification based on high-level and low-level network configuration, signaling-plane and user-plane data.
├── TS3/test.json          # Scenario 3: root cause remediation based on high-level and low-level network configuration, signaling-plane and user-plane data.
├── requirements.txt       # Dependencies
├── README.md              # This file

Main dependencies include:

fastapi
uvicorn
fastmcp
pandas
requests
openai
numpy
tqdm

Running the Benchmark Environment

Option A — FastAPI Server (HTTP Tools)

export TELELOGS_AGENT_CONFIG="TS1"; python fastapi_server.py

Server address:

http://localhost:7861

Scenario context is managed using the HTTP header:

X-Scenario-Id: <scenario_id>

Available endpoints include:

/scenario
/signaling-plane-event-log (only available in scenario TS1 & TS2)
/throughput-logs
/cell-info
/gnodeb-location
/user-location
/user-speed
/serving-cell-pci
/serving-cell-rsrp
/serving-cell-sinr
/rbs-allocated-to-user
/neighboring-cells-pci
/neighboring-cell-rsrp
/beam-scenario-info
/tools

Option B — FastMCP Server

python fastmcp_server.py

MCP endpoint:

http://localhost:7860

Advantages of FastMCP

Native MCP protocol
Session-scoped scenario context
Cleaner agent logic
Seamless integration with MCP-compatible agents

The FastMCP server exposes the same logical tools as the FastAPI server.

Running the Agent Evaluation

The evaluation script supports only the FastAPI backend. Adapting to FastMCP is however straighforward.

Using FastAPI Tools

export TELELOGS_AGENT_API_KEY=xxxx
python benchmark.py \
  --server_url http://localhost:7860 \
  --model_url http://localhost:7865/v1 \
  --model_name qwen8B \
  --num_attempts 4 \
  --max_samples 20 \
  --save_dir ./results

Evaluation and Scoring

Agents are evaluated along multiple dimensions:

Task Success – Correct root cause identification
Tool Call Efficiency – Average accuracy per number of tool calls
Tool Call Failure Rate
Average number of iterations per task

Citation

If you use TeleLogsAgent in your research, please cite:

@article{Sana2026TeleLogsAgent,
  title={{TeleLogsAgent: A Benchmark for LLM Tool-Use in 5G Network Root Cause Analysis}},
  author={Mohamed Sana and Nicola Piovesan and Antonio De Domenico and Fadhel Ayed},
  year={2026},
  eprint={arXiv:2506.10674},
  url={https://arxiv.org/abs/2506.10674}
}

Downloads last month: 2

Paper:

arxiv.org

Paper for netop/TeleLogsAgent

TeleMath: A Benchmark for Large Language Models in Telecom Mathematical Problem Solving

Paper • 2506.10674 • Published Jun 12, 2025 • 3