I’m Mikhail (Misha) Evtikhiev, a Senior ML Researcher at JetBrains Research in Paphos, Cyprus. I lead projects on post-training of code LLMs. The question behind most of my work: did post-training actually make the model better, and how would we know beyond headline benchmark numbers?

Research

• Post-training and transfer. When do gains from SFT/RL on one task family transfer to others? LoRA vs. full fine-tuning, multi-task and continual learning.
• Evaluation methodology. Construct validity of coding benchmarks, task taxonomies, benchmark creation, designing evaluation.

Selected work

• Out of the BLEU: How should we assess quality of the Code Generation models? (JSS 2023)
• Kotlin ML Pack: Technical Report (arXiv:2405.19250)
• Don’t Claim Benchmark-Oriented Optimization Improves General Coding Capability — Diverse Evaluation Is Required (DL4Code workshop @ ICML 2026, arXiv:2608.13566)

All publications are on the Publications page.

Background

PhD in theoretical physics at the Weizmann Institute (advisor: Ofer Aharony), defended in 2020. Thesis name: “On superconformal field theories and little string theories”. In 2021–2023 I also did mixed-methods software-engineering research on team collaboration.