Tomer Ashuach
Researcher at Technion – Israel Institute of Technology
Technion – Israel Institute of Technology
I’m a PhD student at the Technion, advised by Prof Yonatan Belinkov. My research focuses on the interpretability of language models, with particular emphasis on uncovering their internal mechanisms and understanding how knowledge is acquired and can be unlearned.
Research Interests
- Interpretability in LLMs
- Knowledge and Unlearning in LLMs
- AI Safety and Alignment
news
| Jul 2026 | Presented two oral papers (CRISP and Masked by Consensus) at ACL 2026. |
|---|---|
| Jun 2026 | Invited Talks Gave invited talks at MIT (Jacob Andreas), Brown (Ellie Pavlick), the Kempner Institute at Harvard (PRISM reading group), Boston University (Aaron Mueller), and Northeastern (David Bau). |
| Apr 2026 | Gave an invited talk at the Citadel GQS Colloquium. |
| Jul 2024 | Pre-Doctoral Research School The Cornell, Maryland, Max Planck Pre-doctoral Research School in CS (Jul. 2024) Selected as one of 90 top international students for a prestigious program (https://cmmrs.mpi-sws.org) featuring expert lectures and research mentorship from participating institutions. |
Publications
2026
- ACL 2026 (Oral)
In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics: ACL 2026, Dec 2026We explore whether large language models possess privileged internal knowledge about their own correctness, analogous to human introspection. By evaluating models on conflicting predictions, we discover that LLMs exhibit a domain-specific intuition for factual knowledge tasks that external observers cannot access. - ACL 2026 (Oral)
In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics: ACL 2026, Jan 2026We introduce CRISP, a parameter-efficient method leveraging sparse autoencoders to achieve persistent concept unlearning in LLMs. By automatically identifying and suppressing salient features, CRISP successfully removes harmful knowledge while maintaining model utility.
2025
- ACL Findings 2025
In Findings of the Association for Computational Linguistics: ACL 2025, May 2025We propose REVS, a non-gradient-based method that unlearns sensitive information from language models by identifying and modifying a small subset of relevant neurons. REVS achieves robust unlearning and resists extraction attacks while preserving the model’s underlying integrity.