# Nikita Pavlichenko

> Research engineer working on LLM pre-training: data, ablations and training infrastructure. Member of Technical Staff on Cohere's Foundations team; previously led Mellum, JetBrains' open code LLMs.

Location: Berlin, Germany. Contact: nikita.v.pavlichenko@gmail.com.

## Experience

- Cohere: Member of Technical Staff, Foundations (Nov 2026 – now).
- JetBrains: Staff Research Engineer / Team Lead (Mar 2024 – Oct 2026). Led Mellum, JetBrains' open-source code LLM family, and a team of 7 researchers and engineers.
- Yandex · Toloka AI: AI Research Scientist (Sep 2020 – Mar 2024). Led deep-learning research for Toloka AI: 6 papers on NLP, computer vision and generative AI at NeurIPS, SIGIR and HCOMP.
- MIPT & Inria NANO-D: Research Intern (Mar 2020 – Sep 2020). Developed S-GCN, spherical convolutions on molecular graphs for protein model quality assessment. State of the art on the CASP MQA challenge; published in Machine Learning: Science and Technology.
- Yandex: Software Engineering / ML Intern (2018 – 2019). Built a model for ranking video clips in web-crawler schedules that removed over 90% of deleted videos from search results.

## Selected work

- Mellum2 (2026): A 12B mixture-of-experts model with 2.5B active parameters. I led pre-training: data mixture, filtering, dedup and FIM ablations, attention design for latency, and the runs on 512×H200. https://arxiv.org/abs/2605.31268
- Mellum in the IDE (2025): In-IDE completion with multi-file project context. Competitive with commercial solutions in offline evals and online A/B tests, and used by hundreds of thousands of developers every month. https://arxiv.org/abs/2510.05788
- RLHF at Toloka (2023): Led SFT and RLHF for Toloka's LLM, from data collection to training. The model beat Falcon-40B-Instruct with a >65% win rate on real user requests. Authored the ICML 2023 tutorial on RLHF.
- Crowd-Kit (2021): A Python toolkit for aggregating crowdsourced annotations, alongside benchmark datasets for crowdsourced audio transcription (NeurIPS) and pairwise comparisons. https://arxiv.org/abs/2109.08584

## Publications

- Mellum2 Technical Report. M. Kojic, I. Bondyrev, A. de Moor, J. Shtok, P. Borovlev, K. Lysaniuk, M. Kannan, I. Dolgov, N. Pavlichenko. arXiv:2605.31268, 2026. https://arxiv.org/abs/2605.31268
- Mellum: Production-Grade in-IDE Contextual Code Completion with Multi-File Project Understanding. N. Pavlichenko, I. Nazarov, I. Dolgov, E. Garanina, D. Ustalov, I. Bondyrev, K. Lysaniuk, E. Vu, K. Chekmenev, J. Shtok, Y. Golubev, A. Semenkin, U. Sazanovich. arXiv:2510.05788, 2025. https://arxiv.org/abs/2510.05788
- Best Prompts for Text-to-Image Models and How to Find Them. N. Pavlichenko, D. Ustalov. SIGIR 2023, 2023. https://arxiv.org/abs/2209.11711
- WSDM Cup 2023 Challenge on Visual Question Answering. D. Ustalov, N. Pavlichenko, D. Likhobaba, A. Smirnova. WSDM Cup 2023, 2023.
- CrowdSpeech and Vox DIY: Benchmark Dataset for Crowdsourced Audio Transcription. N. Pavlichenko, I. Stelmakh, D. Ustalov. NeurIPS 2021 Datasets and Benchmarks, 2021. https://arxiv.org/abs/2107.01091
- Spherical Convolutions on Molecular Graphs for Protein Model Quality Assessment. I. Igashov, N. Pavlichenko, S. Grudinin. Machine Learning: Science and Technology, 2021. https://iopscience.iop.org/article/10.1088/2632-2153/abf856
- IMDB-WIKI-SbS: An Evaluation Dataset for Crowdsourced Pairwise Comparisons. N. Pavlichenko, D. Ustalov. NeurIPS 2021 Data-Centric AI Workshop, 2021. https://arxiv.org/abs/2110.14990
- Learning from Crowds with Crowd-Kit. D. Ustalov, N. Pavlichenko, B. Tseitlin. arXiv:2109.08584, 2021. https://arxiv.org/abs/2109.08584
- A General-Purpose Crowdsourcing Computational Quality Control Toolkit for Python. D. Ustalov, N. Pavlichenko, V. Losev, I. Giliazev, E. Tulin. HCOMP 2021 Works-in-Progress & Demos, 2021.
- VLDB 2021 Crowd Science Challenge on Aggregating Crowdsourced Audio Transcriptions. D. Ustalov, N. Pavlichenko, I. Stelmakh, D. Kuznetsov. Crowd Science Workshop at VLDB 2021, 2021.
- Toloka at LMNL Challenge: Learning from Noisy Labels with Annotator Reliabilities and Self-Supervised Training. B. Tseytlin, N. Pavlichenko, D. Likhobaba, A. Smirnova, D. Ustalov. LMNL Challenge.

## Writing

- [I Tried Every LLM Architecture Change since 2020](https://www.nikita-pavlichenko.com/blog/every-llm-architecture-change): The best architecture change since GPT-3 is fine-grained MoE, linear attention and hyper-connections work.

## Links

- github: https://github.com/pilot7747
- scholar: https://scholar.google.com/citations?user=2mSLYhcAAAAJ
- linkedin: https://www.linkedin.com/in/nikita-pavlichenko/
- huggingface: https://huggingface.co/JetBrains