Alba (Ruiran) Su
DPhil Researcher & AI Policy Fellow
University of Oxford · Institute for AI Policy and Strategy
About Me
I am a DPhil student in Engineering Science at the University of Oxford, supervised by Prof. Janet B. Pierrehumbert, and an AI Policy Fellow at the Institute for AI Policy and Strategy (IAPS). My work bridges technical AI research and international AI governance.
Current research: what safety determinations rest on
Frontier safety frameworks govern model releases by defining capability thresholds, so every release decision depends on the claim that a model has not crossed one. My IAPS working paper asks what that claim now rests on, once the benchmarks beneath it stop telling models apart. I classify every capability threshold across four frameworks, from Anthropic, OpenAI, Google DeepMind, and the Shanghai AI Lab with Concordia AI, according to whether the determination rests on a benchmark that still discriminates, a structured human study, or a qualitative judgement. The findings are then mapped onto the EU General-Purpose AI Code of Practice, which makes a legal obligation turn on a capability tier being both measurable and unreached.
A second strand of the work looks at how evaluation evidence transfers across jurisdictions and languages, drawing on primary Chinese sources rather than second-hand summaries.
Technical research
I build causal and multimodal AI frameworks integrating satellite imagery, scientific charts, and text for environmental accountability, including the CLIMATEVIZ benchmark for statistical reasoning and fact verification in high-stakes scientific domains.
Research Interests
- 🛡️ Frontier AI evaluation and dangerous-capability assessment
- 🌐 International AI governance and cross-jurisdictional evidence transfer
- 🔗 Causal inference and explainability (DAGs, ICP, do-calculus)
- 🖼️ Multimodal AI and vision-language models
- 🔬 Argument mining, fact verification, and NLP for scientific discovery
- 🤖 Graph neural networks for social network analysis
Recent News
Publications
Working Papers
Conference & Journal Papers
Book Chapters
Research Projects
The Evidence Behind Frontier AI Safety Determinations Current
Safety frameworks turn evaluations into release decisions, and regulation is beginning to turn those determinations into legal facts. This project classifies every capability threshold across four frontier safety frameworks by what the not-crossed determination actually rests on, and traces what happens as the benchmarks beneath those thresholds saturate. The findings are mapped onto the EU General-Purpose AI Code of Practice, and compared across five jurisdictions spanning statutory enforcement, pre-deployment testing without statutory power, state-institutional benchmarking, and published multilingual methodology.
Methods: primary-source document analysis, a published coding protocol for evidence classification, cross-framework comparison, expert interviews.
CLIMATEVIZ: Benchmark for Scientific Chart Reasoning
A large-scale benchmark of 49,862 chart-claim pairs and 2,896 expert-curated charts, sourced from NOAA, the UK Met Office, and Copernicus, designed to evaluate statistical reasoning and fact-checking on scientific charts. It bridges visual and textual evidence to benchmark vision-language models in high-stakes scientific domains.
Causal Graph Discovery for Scientific Claims
Graph-based causal discovery using invariant causal prediction to detect spurious correlations and validate scientific claims across distributional shifts. The same machinery transfers to verifying evaluation claims across jurisdictions.
Methods: NOTEARS, PC, FCI, GES; do-calculus; mediation analysis; counterfactual reasoning.
Climate Discourse Analysis via Graph Neural Networks
Graph attention networks modelling 1,397 scientific entities to study climate discourse and misinformation on social media, reaching 79% accuracy in detecting climate misinformation.
Teaching
University of Oxford (2022 – Present)
Graduate Teaching Assistant, Department of Engineering Science
- Computational & Corpus Linguistics: tutorials on statistical analysis, graphical models, and causal reasoning in linguistic data.
- Cybersecurity: tutorials on threat analysis and vulnerability assessment.
- NLP for Mental Health: statistical methods for observational data, with an emphasis on confounding control.
- AI for Education: supervised projects on intervention design and the evaluation of AI-based tools.
Professional Experience
IAPS – AI Policy Fellow
2026 | Supervised by Miro Pluckebaum and Clement Neo
- Authored a working paper on the evidence base for frontier AI safety determinations, described under Projects.
- Conducted structured interviews with evaluation practitioners and policy staff across the UK, EU, and China.
Climatic Change (Springer) – Editorial Reviewer
2026 | Special collection: NLP and AI as Climate Solutions
- Reviewing submissions on the opportunities and limits of NLP and AI as climate solutions.
ACL 2025 – ClimateNLP Workshop Co-Organizer
2025 | Vienna, Austria
- Coordinated peer review and facilitated interdisciplinary discussion on responsible AI for high-stakes applications.
China Telecom Research Institute – AI Technology Researcher
2021 – 2025 | Beijing, China
- Built causal models for complex systems using structural equation modelling and knowledge graph reasoning.
- Applied graphical causal inference and retrieval-augmented generation to automated decision support; presented at Linux Foundation conferences.
- Published several IEEE conference papers on intent-based networking and 6G systems.
Tsinghua University – Assistant Researcher, Smart City Program
2020 – 2021 | Beijing, China
- Investigated mechanisms of regional inequality using observational data and mediation analysis.
Awards & Honors
Full scholarship awarded for academic merit and leadership potential.
Research on frontier AI evaluation and how it feeds framework-level and regulatory decisions.
National award recognising academic achievement by Chinese students studying overseas.
Two-time recipient, for causal verification frameworks and statistical reasoning in high-stakes AI.
Awarded for research on interpretability and safety metrics of large-scale foundation models.
Recognised for cross-disciplinary leadership in responsible AI.
Competitive grant supporting Sino-British academic exchange and international research collaboration.
Departmental honour for academic standing in Computer Science.
Beyond Research
🎵 Music & Composition
I play piano, keyboards, guitar, ukulele, Irish flute, harmonica, handpan, and dizi. I compose original music blending folk traditions with neo-classical, indie pop, tropical house, and EDM, and publish on platforms including NetEase.
🎼 Listen
🎧 Winterbird🎨 Creative Arts & Writing
I am a member of the China Writers Association, writing fiction, poetry, and environmental essays, and a member of Procreate Artists, working in digital art alongside acrylic and watercolour painting, often on nature-inspired themes.
