Optimization Theory
Convergence analysis for minimax and LMO-based methods, including Frank-Wolfe variants and optimizers used in deep learning.
Igor Ignashin
ML researcher working on optimization and LLM pipelines
Alma mater: Moscow Institute of Physics and Technology (MIPT).
My work is split between optimization research and practical ML systems. On the research side, I work on SGD dynamics, minimax and LMO-based optimization, and traffic-assignment models. On the systems side, I build training and evaluation pipelines for LLM post-training and inference experiments.
Previously, at BRAIn Lab / MIRAI under the supervision of Alexander Beznosikov, I worked on teacher-generated math data for LLM post-training, reward parsing, SFT/KL-distillation experiments, benchmark reporting, and offline evaluation of early-exit/adaptive decoding ideas in vLLM. I also collaborate with Demyan Yarmoshik at LAB MMO in Alexander Gasnikov's group. I recently also worked as a visiting research student at MBZUAI under Eduard Gorbunov.
I also collaborate on SGD analysis and multi-agent reinforcement learning with Andrei Leonidov's team.
Research Focus
Optimization, stochastic dynamics, and practical ML systems
Stochastic Dynamics
Experiments and theory for finite-step SGD dynamics beyond Brownian-motion approximations and standard Langevin models.
LLM Efficiency
Post-training and efficiency projects on SFT, teacher distillation, pruning, early exit, multi-agent RL, and LLM training dynamics.
Selected Publications
Recent papers and preprints
Aligning Distributionally Robust Optimization with Practical Deep Learning Needs
Accepted to a NeurIPS 2025 workshop.
Why SGD is not Brownian Motion: A New Perspective on Stochastic Dynamics
NeurIPS 2026 submission on stochastic dynamics of SGD.
Frank-Wolfe Modifications for Equilibrium Traffic Assignment
Published in Computer Research and Modeling.
Stochastic Origin Frank-Wolfe for Traffic Assignment
Presented at TFN-2025 and accepted to the Journal of Mathematical Sciences, Series B.
Modeling Skiers Flows via Wardrop Equilibrium in Closed Capacitated Networks
Presented at TFN-2025 and accepted to the Journal of Mathematical Sciences, Series B.
Conjugate Frank-Wolfe in Machine Learning
Presented at OPTIMA and accepted to CCIS.
Talks and Media
Public talks, programs, and media mentions
Efficient approaches to compressing large language models
Public lecture on distillation, structured pruning, and early exit for large language models.
LLM compression at the AIRI summer school in Tomsk
RIA Tomsk quoted my explanation of layer removal for making large language models smaller while keeping useful quality.
Optimization dynamics and traffic-flow papers
Habr research reviews highlighted my work on SGD dynamics and traffic-flow optimization.
Stochastic Origin Frank-Wolfe for Traffic Assignment
Presentation at the Traffic Flows on Networks conference at the Sirius Mathematics Center.
Projects
My work and team research
My projects
Research and engineering projects where I am the main author or a direct contributor.
BRAIn Lab team projects
Projects I have led or supervised with student teams; public links are shown once the repository is ready.