Melis Ilayda Bal

^{^{mbal [at] tuebingen [dot] mpg [dot] de}}

Hello!

I am a third-year PhD student jointly affiliated with the Max Planck Institute for Intelligent Systems, in Tübingen, Germany, and École Polytechnique Fédérale de Lausanne (EPFL).

I conduct my research within the Learning and Dynamical Systems (LDS) group at MPI-IS and LIONS lab at EPFL, under the supervision of Michael Muehlebach and Volkan Cevher. My PhD is supported by the Amazon-MPI Science Hub fellowship. I hold an M.Sc. in Operations Research and a B.Sc. in Industrial Engineering, with a minor in Computer Engineering, from Middle East Technical University (METU), Turkey.

My research focuses on robust and efficient training of foundation models, specifically aimed at developing optimization and game-theoretic methods for risk-aware, modular, and efficient training of large language models and multi-agent systems. Beyond this, I have broad interests including algorithmic decision-making, reinforcement learning and online learning.

– I’m open to collaboration via an internship/MSc thesis project at the LDS group. Feel free to check our website & reach out via email!

– I’m also looking for a PhD research internship opportunity in 2026/2027.

news

Feb, 2026	New preprint is out: SALAAD: Sparse And Low-Rank Adaptation via ADMM for Large Language Model Inference
Oct, 2025	I gave a talk: “Train on Less, Learn More: Risk-Aware Input Selection for Efficient & Robust LLM Training” at the 𝗔𝗜+𝗫 𝗦𝘂𝗺𝗺𝗶𝘁 𝟮𝟬𝟮𝟱, Zurich AI Festival in the Efficient LLMs Finetuning track. ✨
Jul, 2025	I have received honorary mention for the MPI-IS Outstanding Female Doctoral Student Prize!
May, 2025	New preprint is out: ESLM: Risk-Averse Selective Language Modeling for Efficient Pretraining
Jan, 2025	ICLR 2025: Two papers accepted! I will present our work on Adversarial Training for Defense Against Label Poisoning Attacks and Optimistic Games for Combinatorial Bayesian Optimization with Application to Protein Design, see you in Singapore!
Jan, 2024	Passed my PhD Qualifying Exam! Thanks to committee members Dr. Michael Muehlebach, Prof. Andreas Krause and Dr. Manuel Gomez Rodriguez. ✨
Dec, 2023	Joined Learning and Dynamical Systems Group (LDS) at MPI Intelligent Systems led by Dr. Michael Muehlebach.
Mar, 2023	Joined Learning and Adaptive Systems (LAS) - Interactive Learning Group at ETH Zurich & MPI Intelligent Systems led by Prof. Andreas Krause.

selected publications

2025

ArXiv

ESLM: Risk-Averse Selective Language Modeling for Efficient Pretraining

Melis Ilayda Bal, Volkan Cevher, and Michael Muehlebach

arXiv:2505.19893 [cs.LG], May 2025

Bib HTML

@article{mib-eslm,
  author = {Bal, Melis Ilayda and Cevher, Volkan and Muehlebach, Michael},
  title = {ESLM: Risk-Averse Selective Language Modeling for Efficient Pretraining},
  journal = {arXiv:2505.19893 [cs.LG]},
  year = {2025},
  month = may,
}

ICLR 2025

Adversarial Training for Defense Against Label Poisoning Attacks

Melis Ilayda Bal, Volkan Cevher, and Michael Muehlebach

In International Conference on Learning Representations, May 2025

Bib HTML

@article{mib-floral,
  author = {Bal, Melis Ilayda and Cevher, Volkan and Muehlebach, Michael},
  title = {Adversarial Training for Defense Against Label Poisoning Attacks},
  journal = {In International Conference on Learning Representations},
  year = {2025},
}

ICLR 2025

Optimistic Games for Combinatorial Bayesian Optimization with Applications to Protein Design

Melis Ilayda Bal, Pier G. Sessa, Mojmir Mutny, and Andreas Krause

In International Conference on Learning Representations, May 2025

Preliminary version presented at NeurIPS 2023 ReALML Workshop, [Best paper award finalist*, Spotlight talk*]

Bib HTML

@article{mib-gameopt,
  author = {Bal, Melis Ilayda and Sessa, Pier G. and Mutny, Mojmir and Krause, Andreas},
  title = {Optimistic Games for Combinatorial Bayesian Optimization with Applications to Protein Design},
  journal = {In International Conference on Learning Representations},
  year = {2025},
  note = {Preliminary version presented at NeurIPS 2023 ReALML Workshop, [Best paper award finalist*, Spotlight talk*]},
}