My (Chiffon) Nguyen

My (Chiffon) Nguyen she/her

Nguyễn Trà My / 阮沐茶 / 윈자미

AI Research for Broader World & Life-long Learning

I research current and future AI that are safe and empowering for more people. I’m interested in technical AI safety, pluralistic AI and agents, currently concerned about:

  1. auditing AI and building tools for AI governance (GPS-Bench, IncidentBench)
  2. loss of control risks and scalable oversight (CoT monitoring)
  3. dynamics and safety of multi-agent systems (e.g., collusion, secret loyalties)
  4. multilinguality (SEATauBench), multicultural awareness (CulturalRiddles), and pluralistic alignment
  5. AI in epistemic contexts (e.g., persuasion, spread of misinformation)

I’m a SPAR Fellow (Fall ’26), mentored by Dr. Linh Le and David Williams-King (Lida Safety), researching AI policy simulation. I also contract for Lida working on multi-agent safety, and for Equistamp. I’m affiliated with SEACrowd and EvalEval Coalition.

Previously I worked with Dr. Samuel Cahyawijaya and Dr. Julia Kreutzer (Cohere Labs), and Eyon Jang (then MATS 8 now Scale AI).

I am open to research collaboration and best reached by email.

Latest updates

see all →

Selected publications

see all →

CulturalRiddles: A Multicultural Benchmark of Riddles

My Chiffon Nguyen*, Tianyi Hu*, Vinod Anbalagan*, Henry Gagnier*, Pouya Sadeghi*, Farah Abdou, , Károly Boczka, Julia Kreutzer
EMNLP Workshop MRL Extended Abstract & Preprint • 2026
Abstract
As LLMs expand their language coverage, evaluating how well they understand and represent diverse cultures becomes increasingly important. However, many benchmarks remain English-centric, with limited cultural coverage and a focus on factual knowledge over its application in real-world cultural contexts. We propose CulturalRiddles, a benchmark of original, culturally grounded riddles created entirely by humans through a participatory community effort, spanning 62 cultures (45 countries, 51 languages). These riddles are deeply rooted in their respective cultures and require abductive and compositional reasoning, alongside detailed factual and social knowledge, language skills, and an understanding of culturally specific associations. We benchmark 28 LLM configurations using both automatic and large-scale human evaluation. Even frontier models achieve at most 22% accuracy in some communities. We publicly release all artifacts alongside an open contribution process to expand the benchmark's linguistic and cultural coverage.
Code Data PDF Thread Website
Evaluation/Benchmarking Multilingual AI Multicultural AI

GPS-Bench: A Governance Policy Simulator for Automating Policy Analysis

Linh Le, Melanie Bui, My Chiffon Nguyen, Zachary Schlosser, David Williams-King
Under review • 2026
Abstract
Policy analysis requires more than predicting whether a proposal will pass: it requires identifying who will be affected, how those actors respond, and what follows. LLM-based policy simulations model these processes at scale, but their validity is hard to establish when plausible behaviour is never compared with observed outcomes. We introduce GPS-Bench, an evidence-grounded benchmark for governance policy simulation that links policies to relevant actors, actor actions and downstream impacts using legislative records, lobbying disclosures, regulatory documents, corporate filings, economic data and other public evidence. Actors are reconstructed from the dated record rather than prompted as archetypes, so a persona is an evidence object with provenance; a human-annotated pool forms the Gold evaluation set, while cases labelled by a separate LLM from retrieved evidence are treated as Silver supervision and never as test labels. Because every inference mode reads the same grounded state and emits the same schema, GPS-Bench turns 'does multi-agent simulation help?' into a controlled comparison: we contrast joint reasoning, independent and communicating actor agents, graph-based methods and weight-level fine-tuning over one policy state. Fine-tuning on the grounded record gives the strongest actor-level impact prediction, and decomposition does not beat it; what decomposition adds is mechanism. Agents hold private, non-identical evidence, each seeing its own exposure clause, and address named partners with concrete joint proposals—what they offer, what they need in return, and why acting together beats acting alone—so the coalitions that form can be checked against the commitments the record holds. GPS-Bench therefore gives a common empirical setting for studying when evidence, actor modelling and multi-agent interaction improve the prediction and interpretation of policy outcomes.
PDF
Evaluation/Benchmarking AI Governance Agent

SEATauBench: Progressively Adapting Tool-Agent-User Evaluation Into Low-Resource Southeast Asian Languages

My Chiffon Nguyen*, Aulia Adila*, Saksorn Ruangtanusak*, Kittiphat Leesombatwathana*, Vissuta Gunawan Lim*, Patomporn Payoungkhamdee, Samuel Cahyawijaya
EMNLP Findings • 2026
Abstract
While AI development and evaluation for Southeast Asia (SEA) has grown rapidly, agent capabilities in regional languages are still poorly understood despite its importance to sovereign AI. To fill this gap, we introduce SEATauBench, the first agent-focused evaluation framework for SEA sovereign AI. It adapts τ²-Bench (Barres et al., 2025) to five languages—Mandarin, Vietnamese, Thai, Indonesian, and Filipino—and evaluates agents cross three progressively localized settings that vary the language of user-agent interaction, tool specifications, and task domains. Across three models, we find that English agent capabilities transfer reasonably well when only the conversation language changes, but quality and robustness degrade sharply as more task contexts are localized, with the largest losses in full domain adaptation. We also highlight the limits of English-only agent assessment for predicting agent capabilities in SEA languages. More broadly, SEATauBench provides a diagnostic benchmark and reusable adaptation pipeline for building reliable multilingual agents for linguistically diverse regions.
Code PDF Website
Evaluation/Benchmarking Agent Multilingual AI

Selected projects

see all →

Studying how monitor effectiveness changes as the capability gap between monitor and target models widens, with a case study on distinguishing sandbagging from genuine incapability.

Python · LLM evaluation · Chain-of-thought monitoring

Career highlights

see CV →

Miscellaneous

  • My Vietnamese name is ‘Trà My’, which is Camellia japonica (tea flower).

  • I attended Minerva University (🇺🇸 AY 21-25) studying Machine Learning & Statistics, with peers from 50+ countries and living in 6 cities as part of its (old) global immersion model: Seoul (South Korea), Taipei, Hyderabad (India), Buenos Aires (Argentina), Berlin (Germany). This experience heavily shapes my worldviews and motivates my research focus in diversity and collaboration.

  • Before that, I went to Foreign Language Specialized School (🇻🇳 Chuyên Ngoại Ngữ, acceptance rate ~11%) during AY 17-20, where I studied English and French.

  • I speak a bit of 🇨🇳 Mandarin Chinese (HSK4/B1) and 🇫🇷 French (A2).

  • I like reading, cooking, tea, cycling, teaching, travel, event organizing, philosophy, and politics.

  • My favorite book series are Dune, Project Hail Mary, and 契子.

  • My friends asked me to collect tech recommendation in a page here.

  • Donate to charities if you can afford it! Some options are World Food Programme, Doctors Without Borders, and Giving What We Can.