How can we build LLM agents whose behaviour is reproducible, verifiable, and robust?
Reliability across changing models, tools, environments, prompts, and execution conditions.
Mathematics × Artificial Intelligence
Measuring what machine intelligence actually does.
I am Bowen Wang, a mathematics-trained AI researcher working at the intersection of LLM agents, AI evaluation, and mathematical methods for machine learning. I build evaluation frameworks that make modern AI systems reproducible, decomposable, and scientifically comparable.
Bristol, UK Open to PhD
opportunities · 2027
Public speech & audio models
Standardised end-to-end in the SURE-EVAL onboarding workflow — one shared protocol for runtime and inference validation.
Bilingual policy paragraphs
Annotated through an LLM-assisted, human-verified pipeline for privacy and trustworthiness research.
Micro-F1 · evidence localisation
Reproducible baselines for locating the exact evidence spans that justify a compliance decision.
Research direction
Evaluation as a structured scientific problem.
My work asks how principled modelling, reproducible execution, and agentic systems can help us diagnose and improve modern multimodal AI — not score it, but explain it.
LLM AgentsAI EvaluationMathematical Methods for AI
Trustworthy AISpeech & Multimodal AI
How can mathematical structure help us design better AI evaluation metrics?
Optimisation, partial orders, and error decomposition that reveal failures hidden by aggregate scores.
How should we evaluate systems with complex temporal, structural, and multimodal dependencies?
From overlapping speech to agent tool trajectories, I study partially ordered, asynchronous, and weakly aligned information across modalities.
“Reproducibility is not a final checkbox. It is the infrastructure that makes evaluation scientifically useful.”
Publications & current work
Research grounded in evidence, not headline scores.
-
Published EECR 2026 · IEEE
A Robust Hybrid Ensemble for Tri-Axial Vibration-Based Running Speed Estimation With Leakage-Free Stacking
A leakage-free stacking ensemble that turns noisy wearable vibration signals into reliable running-speed estimates — with the evaluation protocol designed so no information crosses the train/test boundary.
DOI 10.1109/EECR69522.2026.11549312 ↗ -
Accepted NCMMSC 2026
SURE: Standardized Multimodal Model Onboarding and Reproducible Evaluation Framework
Standardised 18 public speech and audio models across six task families, with reproducible runtime and inference validation — the infrastructure layer this site’s evidence strip is drawn from.
Project repository ↗ -
Submitted ICASSP 2027 · under review
SAER: Decomposed Evaluation Metric for Speaker-Attributed ASR
A time-constrained partial-order alignment that separates lexical recognition errors from speaker-attribution errors, so multi-speaker ASR systems can be diagnosed instead of only ranked.
Manuscript under review
Research experience
From trustworthy data to reproducible multimodal evaluation.
Mar 2026 — Sep 2026
Supervisor: Prof. Kai Yu, IEEE Fellow
Shanghai Jiao Tong University
LLM Agents · AI Evaluation · Speech & Multimodal AI
- Co-developed a reproducible agentic evaluation framework spanning 18 public speech and audio models.
- Designed structured evaluation for speaker-attributed ASR using temporal constraints and exact error decomposition.
Jun 2025 — Mar 2026
Supervisor: Prof. Yan Zhang
Institute of Information Engineering, Chinese Academy of Sciences
LLM-assisted AI · Trustworthy AI · Privacy & Security Analytics
- Built a collaborative LLM annotation workflow covering 28,000+ bilingual privacy-policy paragraphs.
- Developed reproducible baselines and evidence-span localisation reaching Micro-F1 = 0.91.
Selected projects
Systems, metrics, and interpretable machine learning.
Active
Reproducible AI evaluation
SURE-EVAL
A standardized onboarding and evaluation workflow for heterogeneous speech, audio, and multimodal models.
18/18
models passed one-shot onboarding after validation redesign
View integration notes ↗
Manuscript in review
Mathematical evaluation
SAER
A decomposed metric separating lexical recognition errors from speaker-attribution errors in multi-speaker ASR.
7
public SA-ASR pipelines evaluated on AISHELL-4 and AliMeeting
Research manuscript · ICASSP 2027 submission
Completed
Interpretable computer vision
Decoding Feline Pain
Random Forest and ResNet18 pipelines with Grad-CAM and targeted occlusion studies for interpretable facial-region analysis.
2
complementary landmark and full-image learning routes
View code and schema ↗
Education
A mathematical foundation for responsible AI.
MSc Engineering with Artificial Intelligence
University of Bristol · Bristol, UK
BSc Mathematics
University of Bristol · Upper Second-Class Honours
Let’s connect
I am interested in PhD opportunities starting in 2027.
If your group works on AI evaluation, agent reliability, mathematical machine learning, or multimodal systems, I would be glad to hear from you.
wwwency2003@outlook.com