Ahmed Y. Radwan

Research profile

Research summary

Ahmed Y. Radwan is a researcher at the Vector Institute (Canada) working on the evaluation and governance of trustworthy AI systems. Profiled from OpenAlex bibliographic data, the record shows a tightly concentrated agenda: the five most frequent topics — Ethics and Social Impacts of AI (6 works), Explainable AI (5), Topic Modeling (4), NLP Techniques (4), and Multimodal Machine Learning Applications (3) — all orbit one question, how to measure whether AI systems are reliable, fair, and explainable, which yields a perfect focus score of 100. Activity is rising steeply: 13 works in the last three years, with counts by year jumping from 1 in 2024 to 12 in 2026. Two reviewed papers anchor the catalog: an eight-dimension unified framework for evaluating trustworthy LLMs, agentic AI, and multimodal systems (called conceptually strong but empirically unvalidated), and FairLens, a fairness benchmark for vision-language models in high-stakes decisions spanning 8 models that treats abstention as a first-class metric. In both, Radwan holds a middle author position without corresponding status, inside a compact circle of 4 distinct coauthors. Citation indicators remain early-stage (h-index 2, i10-index 1), so this profile is best read as that of an emerging contributor whose output is concentrated in the current year.

Research keywords

Highlights

• Eight-dimension framework unifies LLM, agentic, and multimodal trustworthiness evaluation

  This 2026 arXiv preprint proposes a single evaluation scaffold spanning large language models, agentic AI, and multimodal systems, including score normalization across dimensions. Evolve's review calls it well specified and conceptually strong but empirically unvalidated, and rates reproduction readiness low.

• FairLens makes abstention a first-class fairness metric for vision-language models

  This 2026 benchmark evaluates fairness in 8 vision-language models for high-stakes decision-making, combining demographic parity with abstention-and-soundness measures under an LLM-as-a-judge pipeline. The review values the design but flags missing confidence intervals and significance tests plus an unvalidated judge; reproduction readiness is medium.

• One program: the measurement layer for trustworthy AI

  Topic frequencies — Ethics and Social Impacts of AI (6), Explainable AI (5), Topic Modeling (4), NLP Techniques (4), Multimodal Machine Learning Applications (3) — together with the two 2026 benchmark papers show a single consistent agenda: building frameworks and benchmarks that measure reliability, fairness, and explainability across model types.

Timeline

1. 2024Earliest year in the OpenAlex counts-by-year record, with 1 work published.
2. 2026Publishes A Unified Evaluation Framework for Trustworthy Large Language Models, Agentic AI, and Multimodal Systems on arXiv.
3. 2026Publishes FairLens: Benchmarking Fairness in Vision-Language Models for High-Stakes Decision-Making on arXiv.
4. 2026Annual output in the OpenAlex counts reaches 12 works, up from 1 in 2024, within 13 works over the last three years.

Papers in the Evolve catalog

• 2026-09-18

  A Unified Evaluation Framework for Trustworthy Large Language Models, Agentic AI, and Multimodal Systems

• 2026-09-01

  FairLens: Benchmarking Fairness in Vision-Language Models for High-Stakes Decision-Making

Frequent co-authors