|
Ilia Badanin I am currently pursuing my Master's in Data Science at École Polytechnique Fédérale de Lausanne (EPFL), where I continue to expand my expertise in machine learning. I was fortunate to have an exchange semester at ETH Zurich. My academic journey began with a Bachelor of Science from the Bauman Moscow State Technical University (BMSTU), where I developed a strong foundation in these fields. During my undergraduate studies, I had the opportunity to spend an exchange semester at EPFL, where I completed my Bachelor's thesis under the guidance of Prof. Martin Jaggi. My thesis focused on multiplication-free neural networks, exploring cutting-edge approaches for computational efficiency. Additionally, I contributed to research at CVLab, improving human pose estimation methods and gaining practical experience in advanced computer vision techniques. More recently, I have worked on multilingual large language model interpretability under the mentorship of Chris Wendler, and with the Swiss AI Initiative with Imanol Schlag on the Swiss LLM, Apertus. At ETH Zurich I worked with Xiaozhe Yao on heterogeneous LLM serving (OSDI 2026) and with Alessandro Stolfo on subliminal learning. I contribute to LLM open source: Hugging Face tokenizers and transformers, and GPU kernels. I trained CHSM8, a chess language model that follows natural-language style prompts, from scratch. |
|
News
|
Publications |
|
CHSM8: A Chess Transformer that Follows Natural-Language Style Prompts
Ilia Badanin arXiv preprint, technical report, 2026 arXiv / PDF / Hugging Face A 210M-parameter chess language model that plays in a style described in plain English. Pretrained on 70B moves, then SFT and DPO for text conditioning through a frozen text encoder. Every move is legal by construction, no search. The pure play model reaches over 1900 Elo; open weights, datasets and recipe. |
|
Benchmarking Concept-Spilling Across Languages in LLMs
Ilia Badanin, Daniil Dzenhaliou, Imanol Schlag arXiv preprint, 2026 arXiv / PDF We investigate "language spilling"—a phenomenon where multilingual LLMs exhibit systematic bias toward dominant language representations. We developed a comparative framework that ranks models based on when they resort to dominant-language meanings: stronger models produce more target-language meanings before failing, while weaker models resort to foreign meanings earlier. Results across nine languages and 100 high-polysemy words revealed significant variation in semantic robustness. |
|
OpenTela: Unifying Decentralized Computing Resources for Heterogeneous LLM Serving
Xiaozhe Yao, Youhe Jiang, Ilia Badanin, Qinghao Hu, Robert Matthew Smith, Binhang Yuan, Imanol Schlag, Eiko Yoneki, Ana Klimovic OSDI, 2026 OpenTela is a decentralized serving system that turns existing Slurm-managed HPC clusters into a shared, cross-institutional inference platform. It layers a unified API over diverse serving engines and heterogeneous hardware, using a peer-to-peer gossip network for service discovery and fault-tolerant routing. The heterogeneity-aware scheduler jointly decides model placement and parallelism strategies across mixed GPU types. In continuous operation for over 16 months serving 13 million requests and 15 billion tokens over 142 models to 500+ researchers. |
|
Apertus: Democratizing Open and Compliant LLMs for Global Language Environments
Alejandro Hernández-Cano, ... Ilia Badanin ... Imanol Schlag ACL, 2026 arXiv / PDF The biggest European open-source LLM addressing data compliance and multilingual support. Trained on 15 trillion tokens across 1,800+ languages (40% non-English) with retroactive robots.txt compliance. Available in 8B and 70B parameters. I was responsible for evaluations and post-training. Releases include all scientific artifacts: data preparation scripts, checkpoints, evaluation suites, and training code. |
|
Occlusion Resilient 3D Human Pose Estimation
Soumava Kumar Roy, Ilia Badanin, Sina Honari, Pascal Fua International Conference on 3D Vision, 2024 arXiv / PDF We address occlusion challenges in 3D body pose estimation from single-camera video sequences by representing the deforming body as a spatio-temporal graph. Our refinement network performs graph convolutions to output 3D poses. To ensure robustness, we train with binary masks that disable edges (similar to drop-out techniques), simulating hidden joints for periods of time and training the network to be immune to occlusions. |
Blog
|
Open Source
|
|
Design and source code from Jon Barron |