Ilia Zavidnyi

• Lead research projects on model-to-model communication, inference optimisation, continual and multi-task learning, and reasoning efficiency.
• Design and run post-training experiments for code and reasoning LLMs using supervised fine-tuning and reinforcement learning, including distributed RL on 64 H200 GPUs with NeMo-RL.
• Fine-tuned Mellum, JetBrains’ code-completion model, and designed evaluations on proprietary repositories.
• Developed Step Rejection Fine-Tuning to learn from failed coding-agent runs: 32.2% SWE-bench Verified resolution vs. 30.9% with rejection fine-tuning.
• Built data-generation, training, and evaluation pipelines for software-engineering agents.

• Built ML-backed web-element detection for Aqua’s locator completion, later reused in UI test generation.
• Rebuilt Aqua’s Web Inspector using the Chrome DevTools Protocol; added Cypress support.

I study continual adaptation of vision-language-action (VLA) policies for robotic manipulation: acquiring new skills while retaining existing ones.

I fine-tune π₀.₅ policies using REALM simulation demonstrations and real DROID data, and evaluate in simulation and on hardware. I also research real-time reinforcement learning.

Developed test-time action sampling, filtering, and averaging for π₀-FAST, improving success from 36.5% to 55.5% across eight REALM simulation tasks without retraining.