2026
Position: Text Embeddings Should Capture Implicit Semantics, Not Just Surface Meaning
ICML 2026 · Proc. 43rd Int. Conf. on Machine Learning, PMLR 306 · Seoul, South Korea
01
Profile
database systems · data mining · trustworthy AI · time-series · deployable systems
Anthony K. H. Tung is a Professor at the National University of Singapore, working across database systems, data mining, information retrieval, and trustworthy, deployable AI for operational decision-making.
His current unifying theme is Prudent AI for High-Stakes, Data-Scarce Operational Systems: right-sized, interpretable, auditable, and robust AI for domains where data are scarce, labels are limited, events are rare, and deployment constraints matter.
The program threads interpretable methods into reliable systems and lightweight, deployable tools across text, time-series, tabular, and longitudinal data — connecting methods, systems, and measurable operational outcomes across manufacturing, cybersecurity, and finance/crypto. A second strand views AI as connective infrastructure that elevates collective human potential.
02
Research at a Glance
the Prudent AI framework — from data reality to deployment outcomes
Layer 1
Data reality. Scarce labels, rare events, edge constraints, privacy-sensitive data, and high-stakes human decisions.
Layer 2
Core methods. Interpretable embeddings, diversified retrieval, white-box anomaly detection, controllable generation, and privacy auditing.
Layer 3
Systems & tooling. AI boxes, RAG assistants, agentic workflows, visualization, benchmarking, and operator-facing interfaces.
Layer 4
Deployment outcomes. Manufacturing, cybersecurity, finance, cryptocurrency analytics, and other operational environments.
03
Research Interests & Emerging Themes
seven directions advancing the prudent-AI agenda
i
Prudent AI
Right-sized intelligence aligned to context, cost, and purpose — modular, governable "AI boxes" that lean teams can compose, audit, and deploy.
ii
Just-in-Time AI
On-demand construction of lightweight models from scarce or streaming data for early anomaly and event detection, with controllable time-series generation for rapid adaptation.
iii
White-Box AI
Interpretable embeddings, explainable ranking, and transparent analytics that support contestability, auditing, and trustworthy decision support.
iv
Diversified View / RAG
Retrieval and RAG that surface complementary evidence beyond similarity — diversity-aware re-ranking and interpretable metrics to counter redundancy and echo chambers.
v
Trustworthy Information
Half-truth detection, omission-aware fact verification, and modeling the evolution of misinformation; robustness and safety of LLM reasoning.
vi
Small-Data Event Models
Event-centric modeling under limited labels, rare events, and non-stationarity — combining weak supervision, synthetic data, correlation analysis, and careful evaluation.
vii
DeepConnect
AI as shared connective infrastructure for collective sensemaking and prudent cross-domain execution — knowledge graphs, provenance-linked boundary objects, and governance.
04
Contributions by Pillar
representative systems and named artifacts, organized by methodological pillar
A · White-Box Representation
Interpretable Text Embeddings
CQG-MBQA builds interpretable embeddings from contrastively generated yes/no questions; GSTransform adapts pre-computed embeddings to user instructions via guided space transformation; PRISM produces interpretable political-bias embeddings; QIME extends the idea to ontology-grounded medical text.
B · Diverse Retrieval & Verification
Trustworthy Information
DiversiNews and NEWSCOPE retrieve relevant-yet-diverse news to counter echo chambers; TRACER (with POLITIFACT-HIDDEN) formulates half-truth detection; RADAR performs omission-aware verification via role-anchored multi-agent debate; MPCG models misinformation evolution through persona-conditioned generation.
C · Scarcity-Aware Analytics
Time-Series under Scarcity
CAD detects early anomalies in sensor-based multivariate time series through correlation changes and affected-sensor localization; EADS exposes CAD through an interactive visual system; CTS supports controllable time-series generation for interpolation and extrapolation.
D · Generation & Agents
Benchmarks, Assistants & Workflows
TSGBench standardizes time-series-generation evaluation; CTBench tailors it to cryptocurrency series; TSGAssist uses LLMs and RAG for recommendations; TS-Agent automates financial time-series modeling via structured agentic workflows and reflective feedback.
E · Privacy & Tabular
Privacy, Security & Tabular Analytics
LDSS detects leaked tabular data used in black-box training via local distribution-shifting synthesis; RFOD performs random-forest outlier detection for mixed-type tabular data; related work extends self-explaining, weight-informed analytics to mixed-type tabular clustering.
F · Robustness & Safety
Safety of LLM Reasoning
The Criteria Attack instantiates Reasoning Hijacking — injecting spurious decision criteria that subvert model judgments while leaving the task goal intact — exposing a blind spot in defenses that only detect goal deviation.
05
Deployment & Application Domains
where prudent methods meet operational reality
01 — Manufacturing
Wafer Fabrication
- Virtual metrology & process prediction
- Excursion / drift detection & yield-risk ID
- Rare-regime anomaly detection via synthetic rare states
→ Earlier detection & predictive maintenance with evidence
02 — Cybersecurity
Identity & Access
- Malicious-login & IAM anomaly detection
- Explainable alerts under edge constraints
- Privacy-preserving augmentation & audit
→ Auditable, monitorable, isolatable, recoverable decisions
03 — Finance
Finance & Crypto
- Market & crypto simulation benchmarks
- Time-series forecasting, generation & stress-testing
- Benchmark-driven model selection; LLM workflows
→ Deployment-ready modeling with measurable risk KPIs
06
Current Research Initiatives
active programs & named platforms
AI Prudent Technologies (APT)
A modular ecosystem of lightweight, explainable, and governance-aware AI "boxes" that reduce compute and data demands while improving oversight and trust.
Just-In-Time.AI
Just-in-time construction of lightweight multivariate time-series models for early event detection — controllable synthetic generation for rare conditions, robust augmentation, plug-and-play edge deployment, and conversational operator interfaces.
Explainable Diversified Search in RAG
Extending RAG beyond similarity-only retrieval by balancing relevance, diversity, and user-guided coverage through hybrid retrieval, contextual re-ranking, interpretable embeddings, and visible diversity.
Half-Truth Detection & Omission-Aware Verification
Detecting presented evidence, hidden evidence, inferred claim intent, and the causal impact of missing context — including role-anchored multi-agent debate (RADAR).
Misinformation Evolution Modeling
Persona-conditioned, multi-round generation frameworks simulating how misleading claims evolve through ideological reframing, semantic drift, and audience-specific framing — enabling stress tests for detectors.
Interpretable Representation Learning
Interpretable text embeddings via automatically generated questions, instruction-guided space transformations, and bias-aware cross-encoders that expose human-readable dimensions.
Controllable Generation, Benchmarks & Assistants
Making time-series generation controllable and benchmarkable through CTS, TSGBench, CTBench, TSGAssist, and finance-oriented agentic workflows.
Controllable Unlearning & Privacy Auditing
Selective removal of data via restricted usage policies, data/model capsules, unlearning algorithms, and auditing — with detection of leaked tabular data via synthetic injection and black-box querying.
AI Mediator for Cross-Domain Collaboration (DeepConnect / Synapse)
Cross-domain mediation for experts who use different conceptual languages — using domain corpora and knowledge graphs to establish a shared framework and intervene when miscommunication arises.
07
Selected Publications
2018 – 2026 · filter by topic · verified against DBLP
2026
Reasoning Hijacking: The Fragility of Reasoning Alignment in Large Language Models
ACL 2026 (to appear)
2026
RFOD · Random Forest-based Outlier Detection for Mixed-Type Tabular Data
ICDE 2026 · Montréal, Canada
2026
DeepConnect · A Visual Analytics System for Bridging Interdisciplinary Research Collaborations
IEEE TVCG (to appear) · online demo at deepconnect.sg
2026
IDRBench · Interactive Deep Research Benchmark
CoRR abs/2601.06676
2026
QIME · Constructing Interpretable Medical Text Embeddings via Ontology-Grounded Questions
CoRR abs/2603.01690
2026
Weight-Informed Self-Explaining Clustering for Mixed-Type Tabular Data
CoRR abs/2604.05857
2026
Debating the Unspoken: Role-Anchored Multi-Agent Reasoning for Half-Truth Detection RADAR
CoRR abs/2604.19005
2025
Don't Reinvent the Wheel: Efficient Instruction-Following Text Embedding via Guided Space Transformation GSTransform
ACL 2025 (Long) · pp. 24511–24525
2025
PRISM · A Framework for Producing Interpretable Political Bias Embeddings with a Political-Aware Cross-Encoder
ACL 2025 (Long) · pp. 27719–27733
2025
A General Framework for Producing Interpretable Semantic Text Embeddings CQG-MBQA
ICLR 2025 · Singapore
2025
Uncovering the Bigger Picture: Comprehensive Event Understanding via Diverse News Retrieval NEWSCOPE
EMNLP 2025 · pp. 33939–33957
2025
The Missing Parts: Augmenting Fact Verification with Half-Truth Detection TRACER
EMNLP 2025 · pp. 33979–33996
2025
MPCG · Multi-Round Persona-Conditioned Generation for Modeling the Evolution of Misinformation with LLMs
EMNLP 2025 · pp. 34030–34064
2025
CMS-VAE · A Strategy-Aware Variational AutoEncoder for High-Fidelity Crypto Market Simulation
ICAIF 2025 · pp. 534–542 · Competition Champion
2025
Structured Agentic Workflows for Financial Time-Series Modelling with LLMs and Reflective Feedback TS-Agent
ICAIF 2025 · pp. 924–932
2025
CTBench · Cryptocurrency Time Series Generation Benchmark
NeurIPS 2025 Workshop on Generative AI in Finance
No PDF
2025
Misinformation, Disinformation, and Generative AI: Implications for Perception and Policy
Digital Government: Research and Practice 6(1): 11:1–11:15
2025
FedCare · Towards Interactive Diagnosis of Federated Learning Systems
Frontiers of Computer Science 19(7): 197335
2024
From Zero to Hero: Detecting Leaked Data through Synthetic Data Injection and Model Querying LDSS
PVLDB 17(8): 1898–1910 (VLDB 2024) · Guangzhou, China
2024
TSGAssist · An Interactive Assistant Harnessing LLMs and RAG for Time Series Generation Recommendations and Benchmarking
PVLDB 17(12): 4309–4312 (VLDB 2024, Demo)
2024
DiversiNews · Enriching News Consumption with Relevant yet Diverse News Articles Retrieval
PVLDB 17(12): 4277–4280 (VLDB 2024, Demo)
2024
TSGBench · Time Series Generation Benchmark
PVLDB 17(3): 305–318 · Best Research Paper Award Nomination
2024
EADS · An Early Anomaly Detection System for Sensor-based Multivariate Time Series
ICDE 2024 (Demo) · pp. 5433–5436 · Utrecht, Netherlands
2024
Contextualized Speech Recognition: Rethinking Second-Pass Rescoring with Generative Large Language Models
IJCAI 2024 · pp. 6478–6485
2024
Towards Controllable Time Series Generation CTS
CoRR abs/2403.03698
2023
A New Sparse Data Clustering Method Based on Frequent Items k-FreqItems
SIGMOD 2023 · 1(1), Article 5, pp. 1–28 · Seattle, WA, USA
2023
A Stitch in Time Saves Nine: Enabling Early Anomaly Detection with Correlation Analysis CAD
ICDE 2023 · pp. 1832–1845 · Anaheim, CA, USA
2023
Lightweight-Yet-Efficient: Revitalizing Ball-Tree for Point-to-Hyperplane Nearest Neighbor Search BC-Tree
ICDE 2023 · pp. 436–449 · Anaheim, CA, USA
2023
SAH · Shifting-Aware Asymmetric Hashing for Reverse k Maximum Inner Product Search
AAAI 2023 · 37(4): 4312–4321 · Washington, DC, USA
2023
QALinkPlus · Text Enrichment with QA Data
IEEE Data Engineering Bulletin 47(4): 115–128
2023
SQuAD-SRC · A Dataset for Multi-Accent Spoken Reading Comprehension
IJCAI 2023 · pp. 5206–5214
2018–21
Point-to-Hyperplane Nearest Neighbor Search Beyond the Unit Hypersphere
SIGMOD 2021 · pp. 777–789 · Xi'an, China
2018–21
Locality-Sensitive Hashing Scheme based on Longest Circular Co-Substring
SIGMOD 2020 · pp. 2589–2599 · Portland, OR, USA
2018–21
Sublinear Time Nearest Neighbor Search over Generalized Weighted Space
ICML 2019 · pp. 3773–3781 · Long Beach, CA, USA
2018–21
Accurate and Fast Asymmetric Locality-Sensitive Hashing Scheme for Maximum Inner Product Search
KDD 2018 · pp. 1561–1570 · London, UK
No publications match this topic filter.
Full bibliography & citation counts → DBLP · Google Scholar · ACL Anthology · recent papers (combined PDF)
08
Patents & Applications
filed inventions across time-series generation & anomaly detection
System for Time Series Generation (TSG) Model Selection
International Patent Application
TSGAssist · Interactive Assistant for Time Series Generation Recommendations & Benchmarking
Y. Ang, A. K. H. Tung, Q. Huang, Y. Bao · App. No. 10202402169R
Method and System for Abnormal Signal Detection
A. K. H. Tung, Y. Ang, Q. Huang · PCT/SG2024/050191
CAD · Anomaly Detection for Sensor-based Multivariate Time Series via Correlation Analysis
A. K. H. Tung, Y. Ang, Q. Huang · App. No. 2023103270590
09
Software, Benchmarks & Artifacts
open-source releases · click to view on GitHub
10
Awards & Recognition
competition championships & best-paper honors
Champion · Micron Data Challenge
INFORMS · 2025
Champion · Antalpha Crypto Market Simulation for Risk Estimation Challenge
ACM ICAIF · for CMS-VAE
Best Research Paper Award Nomination · PVLDB / VLDB 2024
for TSGBench: Time Series Generation Benchmark
10-Year Best Paper Award · DASFAA 2015
for Mining Contrast Patterns in Microarray Data
Best Paper Award · DASFAA 2005
for Indexing DNA Sequences Using q-grams
11
Media & Public Engagement
commentary, panels & op-eds on AI, misinformation, cyber resilience & policy
2026
Apr 2026
"AI-generated misinformation targets Singapore"
Expert analysis on cheap deepfake pipelines and the challenges of platform detection and moderation.
Channel 8 News (Focus) · 8worldApr 2026
Op-ed on AI bilingualism — "From using two to beget three to hearing one, knowing ten"
Argues Singapore's three-year "AI-bilingual" push should cultivate a transferable cognitive structure pairing deep domain expertise with AI fluency.
Lianhe Zaobao, Opinion 言论 · zaobao.com.sgMar 2026
Telco service disruptions and national resilience
Called for stronger real-time monitoring and anomaly detection, treating telcos as core national-resilience infrastructure.
CNA Digital · channelnewsasia.comMar 2026
Live studio interview on the Singtel outages
Discussed likely cascading hardware/software/routing failures and the resilience of national communications infrastructure.
CNA · Singapore TonightMar 2026
"Not slow, just don't stand still"
Panellist alongside the Minister of State (MDDI) and Education on AI skills, prompting, self-directed learning, and human purpose.
Channel 8 · Hello Singapore · 8worldFeb 2026
Top-down coordination of AI deployment (Budget 2026)
On the PM-chaired National AI Council and why advanced manufacturing needs precise, sector-specific systems over general-purpose models.
Lianhe Zaobao · zaobao.com.sgFeb 2026
"Hackers lurk within telecommunications systems"
Explainer on advanced persistent threats (APT), stressing monitorable, isolatable, and recoverable systems.
Channel 8 News · 8worldFeb 2026
AI-generated fake videos eroding public trust
On the falling cost of convincing synthetic media — suggesting one effective response is to "counter AI with AI."
Lianhe Zaobao · zaobao.com.sgJan 2026
SG60 technology retrospective
Academic perspective on Singapore's journey from office automation to AI, and AI's potential to unlock collective potential.
Channel 8 News · 8world2025
Aug 2025
Verifying AI outputs
On AI error margins and the need to verify AI-generated answers, including in coding.
Channel 8 · FrontlineAug 2025
SMEs in the AI era
On the data, energy, talent, and cost challenges SMEs face in adopting AI.
Lianhe ZaobaoJun 2025
"The Perfect AI Lover: Sweet Escape or Toxic Trap?"
On AI-generated companions, digital boundaries, and social impact.
Channel 8 · Hello SingaporeMay 2025
Future AI trends
On AI in security, aviation, and port operations, and the role of talent and infrastructure.
Channel 8 · Hello SingaporeApr 2025
"GE2025: Can AI Determine Winners and Losers?"
Panel on AI, elections, public sentiment, and information ecosystems.
Money FM 89.3 · GE Tonight2024 & Earlier
Nov 2024
Exclusive interview on NUS's AI implementation
As AI Lead for Urban Sustainability, on NUS AI adoption and common implementation challenges.
iTnews AsiaSep 2024
"Dark AI" and cybersecurity threats
On the misuse of AI and emerging cyber-threat vectors.
8world · Click ITJun 2024
Deepfake scams
On deepfake-enabled scams and public awareness.
Channel 8 · Hello SingaporeApr 2024
OpenAI Sora
On the implications of text-to-video generation across industries.
Channel 8 · Hello SingaporeFeb 2024
Deepfakes and AI regulation
On deepfake videos and the relationship between technology and legislation.
Lianhe Zaobao12
Service, Teaching & Engagement
leadership, mentoring & partnerships
Professional Service
- Research PC Co-Chair, VLDB 2012 · Vice PC Chair, ICDE 2012
- Workshop Chair, VLDB 2010 · Poster Chair, WWW 2010 · Vice PC Chair, ICDE 2009
- Vice PC Chair (Pre-processing), SIAM SDM 2007 · Co-PC Chair, COMAD 2006
- PC member & reviewer for SIGMOD, VLDB, ICDE, KDD, ICML, ACL/EMNLP, and IEEE TKDE
Teaching & Mentoring
- CS5344 — Big-Data Analytics Technology; graduate & undergraduate modules in database systems, data mining, search & AI systems
- Supervision across Ph.D., Master of Computing & undergraduate research — time-series, trustworthy NLP, retrieval, privacy & AI deployment
- Research leadership over teams on interpretable embeddings, diverse retrieval, fact verification, time-series generation, anomaly detection & financial AI
Industry Partners
MicronSunningDaleStandard Chartered BankGoogleSTT GDCLogicalis
Government Engagement
GovTechCSADSTACSITHTX
Appointments & Leadership
2024 – now
AI Lead, NUS AI Institute (NAII) · Urban Sustainability
2024 – now
Director, Google–NUS Joint Research & Innovation Center
2020 – now
Professor, Dept. of Computer Science, NUS
2019 – 24
NUS Centre on Privacy Technologies (N-CRiPT)
2012 – 18
Deputy Director, SeSaMe · Sensor-enhanced Social Media
2008 – 19
Associate Professor, Dept. of Computer Science, NUS
2001 – 08
Assistant Professor, Dept. of Computer Science, NUS
Education
Ph.D., Computer ScienceSimon Fraser University, Canada
1998 – 2001
M.Sc., Information Systems & Computer ScienceNational University of Singapore
1997 – 1998
B.Sc. (Hons, Second Class Upper), Information Systems & Computer ScienceNational University of Singapore
1996 – 1997
B.Sc., Information Systems & Computer Science (Pass with Merit)National University of Singapore
1993 – 1996
13
On Research & Learning
the heartwares of research
A guiding text
To be constantly learning and practicing what you learn — isn't it the greatest joy of all? To have bosom friends who share your learning from all over the world — doesn't it bring the greatest happiness? To be unknown to people and yet remain undisturbed — isn't that the mark of the most virtuous? — The Analects of Confucius (论语), Chapter 1:1
The world is beautiful because there are similarities and differences. Similarities let us understand one another; differences make each of us unique and interesting.
In human relationships we respect friendship and seniority; in science and mathematics, we respect the truth.
It is more satisfying to help the weak to improve than to make the strong stronger.
The beauty of life comes from having choices among the no-choices — and from what we do with the talent we are given.
Classical voices that guide the work
好学近乎知，力行近乎仁，知耻近乎勇。To love learning is near to wisdom; to act with vigor is near to benevolence; to know shame is near to courage.— Confucius (孔子) · The Doctrine of the Mean
君子和而不同。The noble-minded seek harmony, not uniformity.— Confucius (孔子) · Analects 13:23
言必信，行必果，硁硁然小人哉！True to every word, firm in every deed — unbending as struck stone; a lesser man, perhaps, yet worthy still.— Confucius (孔子) · Analects 13:20
尽信书，则不如无书。To trust every book completely is worse than to have no books at all.— Mencius (孟子) · Book 7B
我善养吾浩然之气。I am skilled at nurturing my vast, flowing spirit.— Mencius (孟子) · Book 2A
反者道之动，弱者道之用。Reversal is the movement of the Dao; yielding is the working of the Dao.— Laozi (老子) · Tao Te Ching 40
知人者智，自知者明。To know others is wisdom; to know oneself is enlightenment.— Laozi (老子) · Tao Te Ching 33
物物而不物于物。Master things — and do not be mastered by them.— Zhuangzi (庄子) · The Mountain Tree
无用之用，方为大用。The use of the useless is the greatest use of all.— Zhuangzi (庄子) · The Human World