Yonatan Belinkov
From Wikipedia, the free encyclopedia
Tel Aviv University (BSc, MA)
Yonatan Belinkov | |
|---|---|
| Alma mater | Massachusetts Institute of Technology (PhD) Tel Aviv University (BSc, MA) |
| Known for | Research on neural language model interpretability, robustness and controllability |
| Scientific career | |
| Fields | Natural language processing, machine learning, artificial intelligence |
| Institutions | Technion – Israel Institute of Technology |
| Website | belinkov |
Yonatan Belinkov is a computer scientist whose research focuses on natural language processing, machine learning and artificial intelligence. He is an associate professor in the Henry and Marilyn Taub Faculty of Computer Science at the Technion – Israel Institute of Technology.[1][2] During the 2025–2026 academic year, he was a visiting scholar at the Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University.[3]
Belinkov is known for work on the analysis and interpretation of neural language models, including methods for probing, understanding and modifying internal representations in neural networks and large language models.[4][5] His research has addressed model interpretability, robustness, controllability, multilingual natural language processing and the behavior of language models under noise or factual intervention.[6]
Belinkov received a BSc and an MA from Tel Aviv University in mathematics and Arabic and Islamic studies. He received a PhD in electrical engineering and computer science from the Massachusetts Institute of Technology in 2018.[3]
After completing his doctorate, Belinkov was a postdoctoral fellow at the Harvard John A. Paulson School of Engineering and Applied Sciences and a postdoctoral associate at the MIT Computer Science and Artificial Intelligence Laboratory.[3] He joined the Technion faculty in 2020.[7] In 2026, the Azrieli Foundation reported that Belinkov, an Azrieli Early Career Faculty Fellow in 2020–2021, had been promoted to associate professor at the Technion's Taub Faculty of Computer Science.[2]
Research
Belinkov's research has examined how neural language systems encode linguistic information and how their internal representations can be analyzed or controlled. In work reported by MIT News in 2017, Belinkov and collaborators studied the layers of neural machine translation and speech-recognition systems, finding that lower layers tended to encode lower-level linguistic information while higher layers captured more abstract properties.[8]
In 2019, MIT News covered work by Belinkov and collaborators on identifying neurons associated with linguistic features in neural machine translation systems. The study described methods for ranking and ablating neurons to analyze the contribution of particular internal units to translation behavior, including phenomena such as number and gender features.[4]
Belinkov has also published on the robustness of neural machine translation. His 2018 paper with Yonatan Bisk, Synthetic and Natural Noise Both Break Neural Machine Translation, argued that neural translation systems could be brittle when faced with synthetic and naturally occurring noise, and evaluated training approaches intended to improve robustness.[9]
A strand of Belinkov's later work has focused on large language models. He was a co-author of Locating and Editing Factual Associations in GPT, a 2022 NeurIPS paper (i.e., "The Rome paper") that proposed methods for identifying and editing factual associations in autoregressive transformer language models.[10] His subsequent research has included work on interpreting hallucinations and internal truthfulness representations in large language models.[11]
In 2024, Belinkov received a European Research Council Starting Grant for the project Control-LM: Controlling Large Language Models, which aims to develop methods for understanding and controlling the internal mechanisms of large language models.[5][12]
Awards and honors
- Harvard Mind, Brain, and Behavior Postdoctoral Fellowship.[6]
- Azrieli Early Career Faculty Fellowship.[2]
- European Research Council Starting Grant, 2024, for Control-LM: Controlling Large Language Models.[5][12]
- Best Paper Award at EMNLP 2024 for Backward Lens: Projecting Language Model Gradients into the Vocabulary Space, co-authored with Shahar Katz, Mor Geva and Lior Wolf.[13]
- Krill Prize for Excellence in Scientific Research, 2025, awarded by the Wolf Foundation.[7]