Beyza Ermis
Also known as Beyza Ermiş
Beyza Ermis is a Senior Research Scientist at Cohere, where her work includes multilingual language models and model evaluation. She joined Amazon Web Services as a Research Scientist in 2017 and coauthored a 2021 method for robust few-shot meta-learning. Her education includes computer engineering at Bilkent and computer science at Boğaziçi, where she completed a doctorate on privacy-preserving data sharing in 2020. At Cohere, she coauthored an LLM evaluation study and Aya Expanse in 2024, multilingual arbitration in 2025, and Tiny Aya in 2026.
Timeline
-
October 2017–November 2022
Robust meta-learning at Amazon
She works as a Research Scientist at Amazon Web Services from October 2017 to November 2022. In 2021, she coauthors a method that selects informative examples to make few-shot meta-learning more robust.
-
2020
Bayesian methods for privacy
In 2020, she completes a computer-science doctorate at Boğaziçi University. Her dissertation explores privacy-preserving data sharing through Bayesian inference, differential privacy, and coupled tensor factorization. She also holds a Master of Science in Computer Science from Boğaziçi and a Bachelor of Science in Computer Engineering from Bilkent.
-
2024
LLM evaluation and Aya Expanse
In 2024, she coauthors Elo Uncovered, which finds that Elo computations can be volatile and that reliability and transitivity can fail in language-model evaluation. In December 2024, she coauthors Aya Expanse, an 8B and 32B multilingual language-model family.
-
July 2025
Multilingual arbitration
At ACL in July 2025, she coauthors Multilingual Arbitration, a method that routes samples through several models based on each model's strengths for individual languages to improve multilingual performance.
-
Since January 2023
Senior Research Scientist at Cohere
She joins Cohere as a Senior Research Scientist in January 2023. At the Cohere Labs Connect Conference in 2025, she speaks as a Research Scientist at Cohere Labs. In 2026, she coauthors Tiny Aya, a small multilingual model family trained on 70 languages.
Career
Education
- 2020Boğaziçi UniversityPhD
- Bilkent UniversityBachelor of Science in Computer Engineering
Most-cited works
- Link prediction in heterogeneous data via generalized coupled tensor factorization
- Memory Efficient Continual Learning with TransformersWith Cédric Archambeau
- Link Prediction via Generalized Coupled Tensor Factorisation
- Linear Bandits with Stochastic Delayed FeedbackWith Claire Vernade
- On the Challenges of Using Black-Box APIs for Toxicity Evaluation in ResearchWith Sara Hooker
Sources and credits
Photo: We have no published photo of Beyza Ermis yet. Send us a free-licensed photo.
- Beyza Ermis - Senior Research Scientist at Cohere | The Org
- The Leaderboard Illusion | Beyza Ermiş, Cohere
- Towards robust episodic meta-learning
- Bayesian approaches for privacy preserving data sharing | Tezler | Tezara
- Elo Uncovered: Robustness and Best Practices in Language Model Evaluation
- Aya Expanse: Combining Research Breakthroughs for a New Multilingual Frontier
- Multilingual Arbitration: Optimizing Data Pools to Accelerate Multilingual Progress
- Connect Conference 2025
- Tiny Aya: Bridging Scale and Multilingual Depth




