How can AI benchmarks become reproducible, auditable, and independently verifiable?
Execution contracts, dependency validation, trusted environments, and attestation across changing models and systems.
Mathematics × Artificial Intelligence
Measuring what machine intelligence actually does.
I am Bowen Wang, a mathematics-trained researcher working across artificial intelligence and mathematical modelling, trustworthy computing, speech and multimodal intelligence, and cybersecurity. I develop mathematically grounded and verifiable AI methods for evaluation, reproducible benchmarking, trusted execution, LLM agents, and statistical risk modelling.
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 mathematical structure, reproducible systems, trusted execution, and agentic methods can make modern AI evaluation auditable — not merely scoring systems, but explaining when and why they fail.
Mathematical & Verifiable AIReproducible EvaluationTrusted Computing
LLM AgentsSpeech & Multimodal AICybersecurity & Risk Modelling
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 model risk and reliability across complex, heterogeneous AI systems?
From overlapping speech and agent execution to privacy evidence and security-risk distributions, I study partially ordered, weakly aligned, and statistically structured information.
“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.
-
Accepted Accepted by NCMMSC 2026.03 — present
SURE: Standardized Multimodal Model Onboarding and Reproducible Evaluation Framework
- Standardised 18 public AI models across six speech and multimodal task families using unified execution contracts, ModelRegistry, dependency verification, and a four-stage validation pipeline — improving one-shot onboarding from 12/18 to 18/18.
- Reproducibility experiments across seven ASR conditions and two TTS subsets identified 0.02–0.52-point protocol-dependent deviations.
-
Submitted Submitted to ICASSP 2027 2026.03 — present
TEA for SURE: Trusted Evaluation through Attestation for Speech Model Benchmarking
- Uses Trusted Execution Environments and remote attestation to cryptographically bind models, data, inference protocols, runtime environments, and evaluation outputs.
- Intel SGX and Gramine experiments produced identical WER under trusted and direct execution while quantifying protected-execution cost; validation spans multiple ASR architectures and system ablations.
-
In preparation Manuscript in preparation 2026.03 — present
SAER: Mathematically Decomposed Evaluation for Speaker-Attributed ASR
- Formulates speaker-attributed ASR evaluation as a time-constrained partial-order dynamic programming problem and exactly decomposes cpCER − CER into speaker-consistency and cross-speaker order-relaxation terms.
- AISHELL-4 and AliMeeting experiments study overlap, timestamp sensitivity, and speaker attribution, exposing hidden failure modes in conventional evaluation metrics.
-
Earlier peer-reviewed work
Published EECR 2026 · IEEEA Robust Hybrid Ensemble for Tri-Axial Vibration-Based Running Speed Estimation With Leakage-Free Stacking
Leakage-free stacking for reliable running-speed estimation from noisy tri-axial wearable vibration signals.
DOI 10.1109/EECR69522.2026.11549312 ↗
Research experience
From trustworthy data to reproducible multimodal evaluation.
Mar 2026 — present
Supervisor: Prof. Kai Yu, IEEE Fellow
Shanghai Jiao Tong University
Reproducible, trusted, and mathematically grounded AI evaluation
- Co-developed SURE, a reproducible evaluation framework spanning 18 public models across six speech and multimodal task families.
- Investigated verifiable benchmarking through Trusted Execution Environments, remote attestation, Intel SGX, and Gramine.
- Modelled speaker-attributed ASR evaluation through time-constrained partial-order dynamic programming and exact error decomposition.
Jun 2025 — Mar 2026
Supervisor: Prof. Yan Zhang
Institute of Information Engineering, Chinese Academy of Sciences
LLM-Assisted Privacy Annotation and Security Risk Modelling
- Designed a heterogeneous LLM workflow combining independent annotation, conflict critique, and human escalation across 28,000+ bilingual paragraphs, 13-way labels, and character-level evidence spans — reducing human intervention to 12.4%.
- Built reproducible BERT-family baselines with evidence-span localisation reaching Micro-F1 = 0.91.
- Combined Zipf/log-normal modelling, path n-grams, domain–path–parameter associations, and graph-based structural mining into interpretable security-risk scores evaluated with ROC/AUC, KS, Precision@K, and Recall.
Selected projects & competitions
Reproducible, trusted, and mathematically valid evaluation.
Accepted by NCMMSC
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 ↗
Submitted to ICASSP 2027
Trusted AI evaluation
TEA for SURE
Verifiable speech-model benchmarking that binds models, data, protocols, runtime environments, and outputs through Trusted Execution Environments and remote attestation.
TEE
Intel SGX and Gramine experiments matched direct-execution WER while measuring protected runtime cost
Remote Attestation · Trusted Computing · Hardware Security
Manuscript in preparation
Mathematically valid evaluation
SAER
A time-constrained partial-order formulation that decomposes lexical recognition and speaker-attribution behaviour in multi-speaker ASR evaluation.
Δ
exact cpCER − CER decomposition tested on AISHELL-4 and AliMeeting
Partial-order DP · Combinatorial Optimisation · Speech Intelligence
Competition · Other project
Decoding Feline Pain
The Aegis Professor Undergraduate Competition image-classification project compares keypoint-based MLP and Random Forest pipelines with a full-image CNN, using systematic occlusion experiments and Grad-CAM to examine the relationship between predictive performance and interpretability.
View code and schema ↗Open-source competition · Tencent Rhino-Bird Program
OpenTenBase AutoVector
Vectorized-execution optimization for OpenTenBase, a PostgreSQL-based distributed HTAP database. All 9 performance targets passed with PASS/ACCEPT verdicts, with portable P95 latency reduced by 56–68% on identical hardware.
View repository ↗Data competition · 2026 UN Data Marathon China
HeatSafe 2030
Urban Heat Action Intelligence: turning climate, urban-environment, and social data into understandable, comparable heat-risk information for city decision-making, presented as an interactive web experience. Advanced to the national Top 12 of the 2026 UN Data Marathon China round.
Visit project site ↗Education
A mathematical foundation for responsible AI.
MSc Engineering with Artificial Intelligence
University of Bristol · Bristol, UK
Foundations of Artificial Intelligence · AI Driven Design and Simulation · AI in Engineering Practice · Large-Scale Data Engineering · Software Development: Programming and Algorithms
BSc Mathematics, 2:1 Class Honours
University of Bristol · Bristol, UK
Statistical Machine Learning · Stochastic Optimisation · Mathematical Programming · Optimisation · Probability and Statistics
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