About

I am a Ph.D. candidate in Information & Communication Engineering at the University of Electronic Science and Technology of China (UESTC), advised by Prof. Yan Zhang (Member of Academia Europaea, the Royal Norwegian Society of Sciences and Letters, and the Norwegian Academy of Technological Sciences). From 2024 to 2025 I was a visiting scholar at the University of Essex, UK, funded by the China Scholarship Council.

My work spans digital twin networks, LLM/VLM benchmarking, deep reinforcement learning, and LLM agent systems. I have published two first-author papers in CAS Q1 journals — my IEEE Internet of Things Journal survey on digital twin networks is both an ESI Highly Cited Paper and an ESI Hot Paper with 1,000+ citations — and I hold one granted Chinese invention patent. I also served as the student lead of a Huawei joint research project, coordinating a seven-person team to deliver a closed-loop digital twin system for edge networks.

Education

Selected Publications & Patents

J1 paper figure
[J1]Yiwen Wu, et al., "Digital Twin Networks: A Survey," IEEE Internet of Things Journal, 2021.
First author ESI Highly Cited ESI Hot Paper 1,000+ citations CAS Q1 · IF 10.238
J2 paper figure
[J2]Yiwen Wu, et al., "Embodied Intelligence-Based Hybrid Edge Computing Networks for Scalable Task Execution in Industrial IoT," Journal of Industrial Information Integration, 2025.
First author CAS Q1 · IF 11.6
J3 paper figure
[J3]D. Pei, Yiwen Wu, et al., "Methodology and Benchmark for Automated Driving Theory Test of Large Language Models," IEEE Transactions on Intelligent Transportation Systems, 2025.
Corresponding author CAS Q2 · IF 8.4
Further work: two first-author papers at IEEE ICCC 2025; papers at IEEE INFOCOM Workshops 2020 (60+ citations) and IEEE WCNCW 2020; one granted Chinese invention patent.

Projects

muxdev — A local LLM agent harness for multi-model coding CLIs2026.02 — 2026.08
Python · Typer · Rich · SQLite · PTY · tmux · Git Worktree · MCP (JSON-RPC) · RAG · LangGraph · pytest. Designed a unified harness that onboards heterogeneous coding CLIs (Codex, Claude Code, Qwen, Kimi) through provider adapters and PTY management — covering session lifecycle, capability probing, output-event parsing and session recovery. Built two execution modes (Direct / Orchestrated) with an assignment DAG for multi-agent decomposition, plus checkpoint/retry for durable, resumable runs. Integrated local RAG, conversation memory, MCP and skill mechanisms; isolated code changes in Git worktrees producing verifiable changesets; and implemented approvals, shell policies, secret protection, rollback and end-to-end observability (SQLite, transcripts, activity timeline).
Huawei Joint Research Project — Real-time high-fidelity digital twin for edge networks · Student Lead2022.04 — 2024.04
Python · PyTorch · time-series forecasting (LSTM / Informer) · deep RL (TD3) · network simulation. Led a seven-person team building a closed-loop "sensing → prediction → twin modeling → decision" system for multi-base-station edge networks. Designed Bi-LSTM-Attention and MHPA-Informer models for joint multi-interface traffic forecasting, cutting MSE by 20% over baselines; fed forecasts into a TD3 agent jointly optimizing user association and transmit power in MU-MIMO, turning scheduling from reactive to foresighted. Delivered real network-management data ingestion, algorithm serving, and a "mapping → policy → evaluation" validation loop.
Horizon Europe MSCA Project — Benchmarking LLM capabilities for automated driving2024.01 — 2025.01
Python · LLM/VLM APIs · prompt engineering · structured-output parsing · multimodal evaluation. Modeled on the UK DVSA driving theory test, built a benchmark of 726 theory questions (617 text + 109 image) and 54 hazard-perception videos with frame-level hazard labels and a time-sensitive scoring scheme. Ran batched inference across mainstream LLMs/VLMs with unified system/user prompts and structured parsing, evaluating accuracy, F1, event-level IoU, throughput, latency and token cost. GPT-4o reached 88.21% overall on driving theory, while hazard perception remains a clear bottleneck for current models.

Honors & Awards

Highlights