About me

I am PhD student at the Language Technology Lab at the University of Amsterdam supervised by Dr. Vlad Niculae. My research is focused on alignment, test-time scaling and self-refinement. I study how LLMs generate outputs probabilistically, and how and when they can self-improve.

News

• Oct 2026 » EMNLP 2026: We extended our Ladders in Chaos self-refinement approach to frontier models! Our UvA MT's participation in WMT26 built on top of weaker models outperforms more advanced models (we achieve top-1 rank in several language directions!).
• Sep 2026 » I attended MT Marathon in Karlsruhe, where I presented a poster about our recent work on self-refinement for MT.
• Sep 2026 » I gave an invited talk at the University of Edinburgh, where I discussed our work on steering and self-refinement.
• Aug 2026 » EMNLP 2026: Our paper Ladders in Chaos: When, How, (and Perhaps Why) Does Test-Time Scaling Improve LLM Machine Translation was accepted to EMNLP 2026! We study when and how test-time scaling actually helps LLM machine translation.
• Nov 2025 » I presented our work at EMNLP UncertaiNLP Workshop at Suzhou! In our paper Asking a Language Model for Diverse Responses, we show that simple non-independent sampling methods can significantly improve the diversity of LLM outputs compared to standard parallel sampling methods.
• Sep 2025 » I presented our work Control the Temperature: Selective Sampling for Diverse and High-Quality LLM Outputs at COLM 2025 in Montreal! We notice that there are “risky” generation timestamps, when sampling is likely to deteriorate the quality. As a solution, we develop an inference time sampling method that dynamically switches between greedy and high temperature sampling depending on the timestamp.
• Aug 2025 » Our paper, On the Low-Rank Parametrization of Reward Models for Controlled Language Generation, was accepted to the TMLR journal! By looking closely at the parameterization of the reward models (q-style vs v-style), we show that q-style reward models have limited rank-capacity, namely, they cannot model all reward functions in context.
• Jun 2025 » I excited to start my internship at Qualcomm AI Amsterdam, where I will work on agentic code fixing diving into multi-turn RL techniques!
• Jul 2024 » I attended LOGML summer (deep learning and geometry) school, where I worked on a project for bayesian inference of parameters of differential equations.
• 5 Dec 2023 » My paper was accepted to the TMLR journal ! We propose a new non-isotropic distribution for probabilistic modeling on Riemannian manifolds.
• Jul 2023 » I spent a great time at the LxMLS 2023 summer school in Lisbon!
• Jun 2023 » I attended OxML summer school (MLx Finance & NLP), which took place at Oxford.
• 8 Dec 2022 » We presented our workshop paper at BlackboxNLP 2022 where we probed deep learning models for source code.
• 1 Dec 2022 » I started my PhD at the Language Technology Lab at the University of Amsterdam supervised by Dr. Vlad Niculae!
• 29 Apr 2022 » We presented our spotlight paper on CodeBPE on subtokenization for source code models at the DL4Code workshop of ICLR 2022 (online)!
• 22 Aug 2021 » Together with Nadezhda Chirkova, we presented our paper Empirical Study of Transformers for Source Code at ESEC/FSE 2021 (online). Here is our code
• 7 Jun 2021 » A Simple Approach for Handling Out-of-Vocabulary Identifiers in Deep Learning for Source Code paper presented at NAACL 2021 with Nadezhda Chirkova

Materials

• Sep 2026 » Invited talk at the University of Edinburgh, on steering and self-refinement with LLMs.