News & Updates
2025 – 2026"Who benefits from AI? Who's left out?" Nov 20, with Dr. Tanya Chung-Tiam-Fook, moderated by Helena Yu.
"Responsible AI Beyond Accuracy: Fairness, Safety, and Sustainability", Evals & Benchmarks track, Nov 6. Event →
Presented Vector's work on agentic AI evaluation, harness engineering, and responsible AI.
Interviewed by Metroland (Loraine Centeno) on spotting deepfake election videos; syndicated across Ontario (InsideHalton, Niagara Falls Review, Waterloo Region Record, OurWindsor). Article →
· Stress-Testing Efficient Responsible-AI Evaluation: When Compute Savings Change Benchmark Conclusions — TAE Workshop arXiv → Project →
· Detecting Deception, Not Deepfakes: Why Media Forensics Needs Social Theories — TAE Workshop arXiv →
· AgentFinVQA: An Auditable, On-Premise Multi-Agent Pipeline for Financial Chart QA — AABA4ET Workshop Project →
· A Unified Evaluation Framework for Trustworthy LLMs, Agentic AI, and Multimodal Systems arXiv →
· Harness-Aware Evaluation of LLM Agents Preprint →
· FairLens: Fairness in Vision-Language Models arXiv →
· Transparency in Agentic AI survey Project →
Quoted on the hidden cost of AI in "AI's Unasked Question: What Did That Cost?" →; deepfake research cited in "Your Face Is Not Yours" →
Survey of generative audio-video deepfakes across the full social media lifecycle: creation, distribution, detection, provenance, and remediation. Preprint →
Consortium feature on our multimodal fairness research; also profiled in the AIXPERT Summer newsletter and "Inside AIXPERT: One Year of Progress". Feature →
Published at the 5th ACM International Workshop on Multimedia AI against Disinformation (MAD '26), Amsterdam. DOI →
Our next-generation bias detection & debiasing toolkit (Qwen3-4B) was presented to Prime Minister Mark Carney. Project →
New German–Canadian partnership for joint AI & ML research, researcher exchanges, and bilateral funding. Presented at HAICON 2026 (Helmholtz Munich, Jun 2026) in the UNLOCK benchmarking session. News →
Position paper arguing that deepfake detection research still targets a late-2010s threat model while intimate-image abuse, voice-clone scams, and emotional fraud dominate real-world harm. arXiv →
Presenting "Sustainable Open-Source AI Requires Tracking the Cumulative Footprint of Derivatives" — introducing the Data and Impact Accounting (DIA) framework. Accepted as a Spotlight Position Paper at ICML 2026. Paper →
Trace-grounded, trajectory-level evaluation that localizes where agentic AI systems break down — where attribution methods cannot. arXiv →
Factuality-aware Direct Preference Optimization — cuts hallucination rates up to 5× (Qwen3-8B) with no auxiliary reward model. Accepted to Findings (2026). arXiv →
AI Governance, Safety and Responsible AI panel alongside IBM Research, Google DeepMind, Microsoft Research, Kush Varshney, and Francesca Rossi. Workshop →
Information Fusion (2026) — Q1, IF 18.6. Project →
AI Open (2026). DOI →
ACM Computing Surveys (2026). DOI →
Education ×2, Workshops ×1, Posters ×3:
· From Data to CO₂: The Hidden Climate Cost of AI — Education Materials ↗
· Prompting Away Stereotypes? Evaluating Bias in Text-to-Image Models — GenProCC Workshop
· Trust, Risk and Security in Agentic AI — LAW@NeurIPS
· Bias in the Picture: Benchmarking VLMs with Social-Cue News Images — Poster
· LinguaMark: Do Multimodal Models Speak Fairly? — Poster
· HumaniBench: Human-Centric Framework for Multimodal Models — Poster
Research Projects
Active & RecentAIXPERT — EU Horizon
Trustworthy AI for industry. Vector Institute selected as 1 of 3 from 137 global applicants. Grant No. 101214389, €515K.
aixpert-project.eu →FairSense-AI
Comprehensive toolkit for fairness, sustainability, and explainability in large-scale AI systems. Bias auditing, emissions reduction, responsible AI compliance.
FairSense-AI →HumaniBench
Human-centric benchmark evaluating fairness, social bias, and empathy in large multimodal models. Multilingual and adversarial test cases. Published in ACM TIST (2026).
ACM TIST →F-DPO
Factuality-aware Direct Preference Optimization — reducing hallucinations in LLMs (up to 5× on Qwen3-8B) with no auxiliary reward model. Accepted at ACL 2026 Findings.
arXiv:2601.03027 →VLDBench
Benchmark for evaluating vision-language models on multimodal disinformation detection with regulatory alignment. Published in Information Fusion (IF 18.6).
VLDBench →SONIC-O1
Fully human-verified benchmark evaluating multimodal LLMs on real-world audio-video understanding — summarization, evidence-grounded MCQ reasoning, and temporal localization across 13 domains, with demographic/fairness analysis. Accepted at EMNLP 2026 (Main).
sonic-o1 →Green AI & Sustainability
Energy footprint benchmarking of AI models, sustainable LLM training, and carbon reduction research. Spotlight at ICML 2026 (Helmholtz AI Workshop).
GreenBench →UnBias-Plus
Next-generation bias detection and debiasing toolkit powered by fine-tuned LLMs (Qwen3-4B). Detects biased language segments, classifies type and severity, suggests neutral rewrites, and returns a fully rewritten unbiased article — all in one structured output.
GitHub →FairSense-AgentiX
Agentic fairness and AI-risk analysis platform using ReAct reasoning loops with dynamic tool selection and self-critique. Multi-modal bias detection across text, images, and datasets. Production-ready FastAPI + WebSocket streaming + React UI. Funded by EU Horizon AIXPERT. JOSS paper under review.
GitHub →Model Immunization
Techniques for immunizing LLMs against falsehoods and harmful content, improving robustness and factuality. Accepted at IEEE WCCI/IJCNN 2026.
arXiv:2505.17870 →Agentic Evaluation
Evaluating and Regulating Agentic AI (Information Fusion 2026), From Features to Actions (trajectory-level explainability), Harness-Aware Evaluation, a Unified Evaluation Framework for trustworthy LLMs and agents, and the Transparency in Agentic AI survey.
arXiv:2602.06841 →Deepfakes in the Foundation-Model Era
Research program on generative audio-video deepfakes: position paper on misallocated detection effort, ACM MAD '26 timeline paper, lifecycle survey of forensics and distribution, and deception-centred evaluation (NeurIPS 2026 TAE Workshop).
Survey →Sustainable Responsible-AI Evaluation
Does cheaper evaluation (batching, INT8/INT4 quantization, benchmark subsetting) preserve accuracy, bias, and fairness conclusions on BBQ and BBQ-V, and what does it cost in energy, CO₂e, and water. NeurIPS 2026 TAE Workshop.
Project →FairLens
Benchmark for fairness in vision-language models across hiring, legal, and healthcare decision contexts.
arXiv:2609.01691 →Equity in Public Health AI
Embedding fairness and equitable decision-making in AI systems for public health surveillance. CIHR HSI Fellowship ($175K) + AI4PH Award ($20K/yr).
CIHR Fellowship →Selected Publications
Full list on Google Scholar · thousands of citationsShowing recent high-impact work. View all on Google Scholar →
Invited Talks & Keynotes
2023 – 2026Professional Service
Program Committees
- NeurIPS — Main Track + RegML, LAW, SoLaR, GenProCC, Position Paper Workshops
- AAAI · ACL · ICLR · ICML · KDD
- IJCAI · CIKM · ECCV · IEEE Big Data
- AI Governance Workshop @ NeurIPS 2025 — Co-organizer
- AAAI Workshop on Responsible LMs (ReLM 2024) — Co-lead
Editorial Roles
- Springer Discover Computing — Editorial Member (2025–present)
- Elsevier CS Portfolio — Transfer Manager & Scientific Editor (2021–present)
- IEEE Transactions on AI — Review Editor (2023–present)
- Frontiers in Big Data — Review Editor (2022–present)
Journal Reviewing
- Nature · Nature Machine Intelligence
- Lancet Digital Health · Science Robotics
- JAMA Network Open · JAMIA
- BMC Public Health · PLOS ONE
- IEEE Trans. Knowledge & Data Engineering
- Artificial Intelligence in Medicine
Grant Review Committees
- CIHR Doctoral Research Awards — Peer Review Committee Member
- National Killam Program — Peer Reviewer