Shaina Raza

Vector Institute · Canada

Unclaimed · AI-structured profileORCID OpenAlex

Vector Institute researcher unifying trustworthiness, fairness and efficiency evaluation for LLMs and multimodal AI

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-17) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.

Research profile

Output—

Not yet scored: the counts-by-year record is incomplete — only 2024 (2 works) and 2026 (16 works) appear — so a stable productivity trend cannot be computed. This axis would measure publication volume and output trajectory over time.

Impact—

Not yet scored: the citation record is minimal (h-index 1, i10-index 0), too thin to support a meaningful impact estimate. This axis would measure citation-based influence — accumulated citations, h-index growth, and uptake of her benchmarks by later work.

Focus100/100

Scores 100/100, the platform's maximum topical concentration. The OpenAlex topic profile is tightly clustered: "Ethics and Social Impacts of AI" (7 works) and "Explainable Artificial Intelligence (XAI)" (7) lead, followed by "Multimodal Machine Learning Applications" (6), "Topic Modeling" (4), and "Natural Language Processing Techniques" (4). All three reviewed papers target the same problem — how to evaluate trustworthy, fair, and efficient AI — confirming a single coherent agenda.

Collaboration44/100

Scores 44/100, indicating a moderate collaboration footprint. The record lists 8 distinct coauthors, and all eight recur as frequent collaborators: Ahmed Y. Radwan, Ahmed El Kady, Aravind Narayanan, Rehana Riaz, Yani Ioannou, Imran Liaquat, Kathryn Hume, and Vahid Reza Khazaie. Authorship positions — first author on the unified trustworthiness framework and last author on FairLens and the stress-testing study — suggest she leads or supervises compact team projects rather than joining large consortia.

Activity86/100

Scores 86/100, reflecting strongly rising recent output. OpenAlex records 18 works in the recent three-year window, and the counts-by-year series jumps from 2 works in 2024 to 16 in 2026 — an eightfold increase in recorded annual output. All three Evolve-reviewed papers carry 2026 dates and arXiv sources, indicating this activity is current and concentrated in preprint-speed release of evaluation frameworks, fairness benchmarks, and methodology stress-tests.

Reproducibility20/100

Scores 20/100 (reproMean 20, based on 1 assessed paper). Evolve's readiness ratings for the three reviewed papers span the range: the unified trustworthiness framework is rated low — conceptually strong but empirically unvalidated; FairLens is medium — valuable design but no confidence intervals, significance tests, or validated LLM judge; the efficient-evaluation stress test is high. The low current score mainly reflects the framework paper and FairLens's missing statistical rigor.

Research summary

Shaina Raza is a researcher at the Vector Institute (Canada) whose work centers on how trustworthy, fair, and efficient artificial intelligence should be measured. Her OpenAlex record (resolved via ORCID 0000-0003-1061-5845) shows a highly concentrated topic profile: seven works each in ethics and social impacts of AI and explainable AI (XAI), six in multimodal machine learning applications, plus topic modeling and NLP techniques. Activity is strongly front-loaded: 18 works in the recent three-year window, with 16 works recorded in 2026 versus 2 in 2024. Three 2026 arXiv papers reviewed by Evolve define her current agenda. As first author she proposed "A Unified Evaluation Framework for Trustworthy Large Language Models, Agentic AI, and Multimodal Systems", an eight-dimension meta-evaluation framework judged conceptually strong but empirically unvalidated. As senior (last) author she led FairLens, a fairness benchmark for vision-language models that treats abstention as a first-class metric across eight models, and a stress-testing study asking whether compute savings from quantization and batching change responsible-AI benchmark conclusions — the latter rated high in reproducibility readiness. Citation impact remains early-stage (h-index 1, i10-index 0), so influence cannot yet be scored. She typically publishes with a compact circle of eight distinct coauthors, including Ahmed Y. Radwan, Ahmed El Kady, and Kathryn Hume.

Collaboration pattern

The collaboration footprint is compact and stable: 8 distinct coauthors, all recurring as frequent collaborators — Ahmed Y. Radwan, Ahmed El Kady, Aravind Narayanan, Rehana Riaz, Yani Ioannou, Imran Liaquat, Kathryn Hume, and Vahid Reza Khazaie. The pattern of a first-author framework paper plus last-author positions on FairLens and the stress-testing study suggests she anchors small teams in lead or senior roles at the Vector Institute rather than participating in large multi-institution consortia. Collaboration score: 44/100.

Research keywords

trustworthy AI evaluationresponsible AILLM benchmarksmeta-evaluationagentic AImultimodal machine learningvision-language modelsfairness benchmarkingexplainable AI (XAI)natural language processingtopic modeling

Highlights

Unified trustworthiness evaluation for LLMs, agentic AI, and multimodal systems
First-author 2026 framework specifying eight dimensions with score normalization to meta-evaluate LLMs, agentic, and multimodal systems. Evolve's verdict: conceptually strong but empirically unvalidated; reproducibility readiness rated low.
FairLens makes abstention a first-class fairness metric for vision-language models
Last-author benchmark probing demographic parity and abstention/soundness across 8 vision-language models in high-stakes decision scenarios. Evolve values the design but flags missing confidence intervals, no significance tests, and an unvalidated LLM judge; readiness medium.
When compute savings change responsible-AI benchmark conclusions
Last-author stress test of efficient evaluation via quantization, batching, and benchmark reduction, showing cheaper evaluation can flip benchmark conclusions — with scope limited to 3 models, 2 benchmarks, one judge, and one GPU. Reproducibility readiness rated high.

Timeline

2024Earliest recorded output in the snapshot: OpenAlex counts-by-year lists 2 works.
2026Recorded output jumps to 16 works; the recent three-year total stands at 18.
2026First-author arXiv framework unifying trustworthiness evaluation for LLMs, agentic AI, and multimodal systems (Evolve-reviewed; readiness: low).
2026Last-author releases: FairLens vision-language fairness benchmark and the efficient responsible-AI evaluation stress test (readiness: high).

Papers in the Evolve catalog

Frequent co-authors

Ahmed Y. Radwan×2 Ahmed El Kady×1 Aravind Narayanan×1 Rehana Riaz×1 Yani Ioannou×1 Imran Liaquat×1 Kathryn Hume×1 Vahid Reza Khazaie×1

About this profile

This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.

All bibliometrics come from an OpenAlex record (identity: openalex) resolved via ORCID 0000-0003-1061-5845; citation figures (h-index 1, i10-index 0) reflect that source only and may undercount.
Counts-by-year coverage is incomplete: only 2024 (2 works) and 2026 (16 works) appear, so multi-year trends could not be computed and the output and impact axes remain unscored.
All three reviewed papers are 2026 arXiv preprints; peer-reviewed venue status is not captured in this snapshot.
This profile is name-resolved from public bibliographic data; "Shaina Raza" may be shared by multiple authors, so confirm identity against the ORCID before reuse.