cv

PhD Researcher, Post-Training Efficiency

DAP Lab, Columbia University

Advised by Eugene Wu and Kostis Kaffes.

• Studying optimization dynamics in RL post-training (RLVR): benchmarking optimizer behavior and loss-landscape changes relative to pretraining, and their effect on downstream tasks.
• Developing staleness-mitigation methods for asynchronous post-training with parallel rollout generation, co-designing algorithms and systems to maximize GPU throughput.

ML Researcher, Pretraining

Yandex Research

• Led research on asynchronous pipeline parallelism for efficient large-scale LLM pretraining; first-author paper accepted to ICML 2026.
• Designed and analyzed novel algorithms for distributed training systems, improving scalability and hardware utilization.

Teaching Assistant

Central University

Introduction to AI, Scientific Studio, Python for Data Analysis.

• Held consultations, graded assignments, and mentored student research projects.

Teaching Assistant

Moscow Institute of Physics and Technology (MIPT)

Algorithms and Data Structures.

• Held consultations and graded assignments for the Department of Algorithms and Programming Technologies.

ML Engineer, Personalization Quality R&D Group

Yandex

• Engineered a production pipeline integrating Vowpal Wabbit-based online features into a large-scale CatBoost model, achieving a 0.1% uplift in core ranking metrics (AUC, nDCG).
• Implemented a parallelized, MapReduce-style data converter that accelerated data processing by 20x, reducing model training time from hours to minutes.

ML Researcher

Basic Research in AI Laboratory (BRAIn Lab)

• Pioneered memory-efficient zeroth-order optimization methods for LLM fine-tuning, achieving a 50% reduction in memory footprint (ICML Workshop 2025).
• Developed and theoretically analyzed novel stochastic and zeroth-order algorithms for decentralized optimization, validated across multiple domains (UAI 2025, ICLR 2026).
• Derived a full second-order (Hessian) characterization of the Transformer block with implications for scaling laws.

PhD

Columbia University

Computer Science

• SEAS Graduate Student Fellowship.
• Expected graduation: May 2031.

B.S.

Moscow Institute of Physics and Technology (MIPT)

Applied Mathematics and Physics

• Data Science Track (MIPT & Yandex School of Data Analysis).
• ML, DL, RL, CV, NLP, RecSys, Bayesian Statistics, Time Series.

SEAS Graduate Student Fellowship

Columbia University

B. T. Polyak Scholarship

Awarded for contributions to numerical optimization methods

Best Student Paper

Neuroinformatics 2024

For "Shuffling Heuristic in Variational Inequalities: Establishing New Convergence Guarantees".

Increased State Academic Scholarship

Moscow Institute of Physics and Technology (MIPT)

Held 2024-2026.

Sign-SGD is the Golden Gate between Multi-Node and Single-Node Learning

ICLR 2026

D. Medyakov, S. Stanko, G. Molodtsov, P. Zmushko, G. Evseev, E. Petrov, et al.

One-Step Gradient Delay is Not a Barrier for Large-Scale Asynchronous Pipeline Parallel LLM Pretraining

ICML 2026

P. Zmushko*, E. Petrov*, N. Abdullaev, M. Khrushchev, S. Horváth. (* equal contribution)

Closing the Curvature Gap: Full Transformer Hessians and Their Implications for Scaling Laws

Submitted to NeurIPS 2026

E. Petrov, N. Kiselev, V. Meshkov, A. Grabovoy.

Sampling of Semi-Orthogonal Matrices for the Muon Algorithm

AI Journey 2025. Published in Doklady Mathematics

E. Petrov, G. Evseev, A. Antonov, A. Veprikov, N. Bushkov, S. Moiseev, A. Beznosikov.

When Extragradient Meets PAGE: Bridging Two Giants to Boost Variational Inequalities

UAI 2025

G. Molodtsov, V. Parfenov, E. Petrov, E. Grigoriy, D. Medyakov, A. Beznosikov.

Leveraging Coordinate Momentum in SignSGD and Muon: Memory-Optimized Zero-Order LLM Fine-Tuning

ICML Workshop on Tiny Titans, 2025

E. Petrov, G. Evseev, A. Antonov, A. Veprikov, P. Plyusnin, N. Bushkov, et al.

Zero Order Algorithm for Decentralized Optimization Problems

AI Journey 2024. Published in Doklady Mathematics

A. Veprikov, E. Petrov, G. Evseev, A. Beznosikov.

Shuffling Heuristic in Variational Inequalities: Establishing New Convergence Guarantees

ICOMP 2024. Submitted to JOTA (Q1)

D. Medyakov, G. Molodtsov, E. Grigoriy, E. Petrov, A. Beznosikov. Best student paper, Neuroinformatics 2024.

ZO-Library, Lead Developer

Open-source library of state-of-the-art zeroth-order optimization algorithms with a torch.optim-style API, enabling memory-efficient fine-tuning with up to 50% memory reduction.

• Related paper: Leveraging Coordinate Momentum in SignSGD and Muon (ICML Workshop on Tiny Titans, 2025).

Aboba, Maintainer

Python library for A/B testing: statistical tests plus CUPED and linearization via an intuitive API.

Rethinking Memory-Efficient Optimizers for LLM Pre-Training

Systematic study of structured optimizer-state compression vs. state quantization under matched memory budgets; peak-memory and communication analysis of optimizer states under sharding (FSDP-style).

• In submission to ICLR 2027.

Hyperparameter Transfer and Scaling Laws for Asynchronous Pretraining

Theoretical and practical transfer rules (learning rate, momentum, batch size, weight decay) and scaling laws in models and tokens for asynchronous pipeline-parallel training; extends our ICML 2026 results.

• Experiments in progress.

Optimization in RL Post-Training (RLVR)

Benchmarking optimizer behavior, loss-landscape changes, and downstream-task effects when moving from pretraining to RL post-training.

• Early stage.

Asynchronous Post-Training

Mitigating policy staleness from parallel rollout generation; algorithm-systems co-design for maximizing GPU throughput in RL post-training.

• Early stage.