Ido Hakimi

Postdoctoral Fellow
ido.hakimi@ai.ethz.ch
OAT X19.1
+41 76 291 44 79
Linkedln
External Website
My research explores test-time learning, where models dynamically adapt and learn during inference, treating each prompt as an opportunity for real-time improvement. Since compute budgets grow, I focus on methods that scale with inference-time compute to enable more adaptive, context-specific reasoning.

2026
Reinforcement Learning via Self-Distillation J. Hübotter, F. Lübeck, L. Behric, A. Baumann, M. Bagatella, D. Marta, I. Hakimi, I. Shenfeld, T. Kleine Buening, C. Guestrin, A. Krause In International Conference on Machine Learning (ICML), 2026 Best paper award* at Test-Time Updates Workshop and oral presentation at Scaling Post-Training for LLMs Workshop at ICLR [bibtex] [abstract] [pdf]
ActiveUltraFeedback: Efficient Preference Data Generation using Active Learning D. Melikidze, M. Schneider, J. Lam, M. Wertich, I. Hakimi, B. Pásztor, A. Krause In International Conference on Machine Learning (ICML), 2026 [bibtex] [abstract] [pdf] [blog]
2025
Local Mixtures of Experts: Essentially Free Test-Time Training via Model Merging R. Bertolissi, J. Hübotter, I. Hakimi, A. Krause In Proc. Conference on Language Modeling (COLM), 2025 [bibtex] [abstract] [pdf]
Efficiently Learning at Test-Time: Active Fine-Tuning of LLMs J. Hübotter, S. Bongni, I. Hakimi, A. Krause In Proc. International Conference on Learning Representations (ICLR), 2025 Best paper award at Fine-Tuning in Modern Machine Learning Workshop at NeurIPS [bibtex] [abstract] [pdf]