Vector AIXpert: Responsible AI Infrastructure for Fairness, Explainability, and Evaluation¶
Vector Institute's contribution to the AIXpert Project: tools, benchmarks, and research for explainable, accountable, and fair AI.
This project represents the Vector Institute's research contributions to the AIXpert Horizon Europe initiative. It focuses on developing tools, datasets, and evaluation pipelines for fairness-aware generative AI and explainable AI systems.
What we do¶
Vector's contribution to AIXpert spans four core areas:
- Explainable & accountable AI. Tools and benchmarks for interpretability, fairness, and transparency in generative and multimodal AI.
- Trustworthy agentic AI. Transparent, auditable, human-in-the-loop agentic systems with measurable trustworthiness metrics.
- Multimodal evaluation. Benchmarks and datasets for audio-video understanding, vision-language assessment, and fairness across domains and demographics.
- Open, reproducible research. Code, datasets, and documentation shared openly to support governance-ready research.
For the full AIXpert vision, consortium, and funding details, see About.
System Architecture¶
Vector's responsible AI pipeline moves data through five stages: from raw inputs to governed, explainable outputs.
Recent Updates¶
- MARS. Video forgery detector for facial manipulation and fully generated video that reports the frames and regions it used. Project page · Preprint · Code.
- Mozaic. Audio-video deepfake benchmark that labels the face and the voice separately and asks which part is fake, what kind, whose identity, and why. Zero-shot models catch fake video far more often than fake audio. Paper coming. Project page · Code · Dataset.
- Deepfakes in The New York Times. Coverage of AI-generated fakes in midterm memes (12 Nov 2026). Article.
- CIFAR AI Safety Week. Shaina Raza presented deepfake generation and detection, including MARS and Mozaic, at CAISI during CIFAR AI Safety Week (12 Nov 2026, Toronto). Post.
- NVIDIA: Responsible AI & Transparency. Shaina Raza presented DIA as a lightweight transparency layer for the energy, water, and carbon footprint of open models and their derivatives (1:30–1:50 pm). Project.
- MLOps North 2026. Shaina Raza spoke on responsible AI beyond accuracy: fairness, safety, and sustainability, from agent trajectories to safety, reliability, and governance (5–6 Nov 2026, Toronto). Summit · Presenter page.
- Deepfakes in the Niagara Falls Review. A Toronto expert on red flags in deepfake videos. Article.
- FairSense-AgentiX published in JOSS. Agentic fairness and AI-risk platform for text, images, and datasets. Paper · Project page · Code · PyPI.
- SONIC-O1 accepted at EMNLP 2026 (Main). Project page · Paper · Code · Leaderboard.
- DiagFlowBench accepted at EMNLP 2026 (Industry). Paper.
- AgentFinVQA accepted at NeurIPS 2026 AABA4ET. Project page · Preprint · Code.
- Detecting Deception, Not Deepfakes accepted at NeurIPS 2026 TAE. Project page · Preprint.
Related Projects¶
A snapshot of Vector's key contributions within AIXpert. Each project has its own repository, documentation, and quickstart.
-
MARS
Two-path video detector for facial manipulation and fully generated video. A residual path reads spatial and temporal differences next to appearance, and a clip-level gate combines it with a semantic path. Across seven datasets it reaches 91.43% average AUROC, and it reports the frames and regions behind each decision.
Project page · Preprint · Code
-
Mozaic
Audio-video deepfake benchmark with a four-way label: real or fake face, and real or fake voice. It asks which part is fake, what kind of manipulation, whose identity, and why. Zero-shot models catch fake video far more often than fake audio, and a held-out face-swap generator drops the best detector to 78.3%. Paper coming.
Project page · Code · Dataset
-
UnBias-Plus
AI-driven toolkit for bias detection and debiasing in text: biased spans, severity, reasoning, neutral replacements, and a full neutral rewrite for more trustworthy workflows.
Project page · Code · PyPI
-
DIA (Data & Impact Accounting)
Open-source toolkit to track energy, water, and CO₂ impact of open-source AI models and their derivatives, with a public dashboard.
Project page · Code · PyPI · Dashboard
-
FairSense-AgentiX
Agentic platform for fairness and AI-risk analysis on text, images, and datasets, published in JOSS. An agent plans each check, selects tools, critiques the result, and records the steps.
Project page · Paper · Code · PyPI
-
SONIC-O1
Real-world benchmark for evaluating MLLMs on audio-video understanding, with a public leaderboard.
Dataset · Code · Leaderboard
-
SONIC-O1 Multi-Agent
Multi-agent framework for audio-video understanding with chain-of-thought reasoning, self-reflection, and temporal grounding.
-
Explainable Agentic Evaluation Framework
Analyzes reasoning traces and interpretability of agentic AI across static and agentic settings.
-
Factual Preference Alignment (F-DPO)
Factuality-aware preference learning to reduce LLM hallucinations without a separate reward model.
Paper · Project page · Code
-
HumaniBench
Fairness-focused vision-language benchmark evaluating foundation models across human-centric demographics.
-
Agentic Transparency
Survey and framework on interpretability, explainability, and governance of agentic AI systems.
View all papers Projects & quickstarts
Citation¶
If you use any of our tools, datasets, or benchmarks, please cite the relevant work. BibTeX entries are available on each paper's entry in the Papers page.
Have feedback or want to contribute? See the Team section on About and open an issue or pull request.