Maxim Zapolskii

ML Researcher · Reasoning in LLMs

eliornnn@gmail.com Telegram @maximzap GitHub LinkedIn

Research interests

Reasoning in large language models; reinforcement learning for LLMs; formal mathematical reasoning (Lean 4); interpretability; mathematical foundations of deep learning; topological and information-theoretic methods in machine learning.

Honours and awards

IMO 2026 — Gold Medal, perfect score (42/42), Huawei AI team · news (AFP)2026
IMO 2025 — Silver Medal (35/42), Huawei AI team2025
AI Mathematical Olympiad (AIMO) 3, Kaggle — Silver MedalSpring 2026
Olympiad “Higher League”, HSE University — 3rd degree diploma (Applied Mathematics and Computer Science)Spring 2021

Experience

Huawei — ML ResearcherApr 2025 – present
Math and code reasoning in LLMs: SFT, RL fine-tuning, test-time scaling and advanced scaffolding; multimodal data, GUI/CLI agents and vision navigation.
  • IMO 2026 — Gold, 42/42 (AFP coverage). Lean 4 formalization of the problems and experiments with the prover model.
  • IMO 2025 — Silver, 35 pts with solutions in Lean 4. Owned problem formalization and trained the SFT checkpoint used for subsequent RL fine-tuning (ToRA). Lean 4 proofs were translated into natural language and submitted to the IMO committee.
  • Long-form reasoning. Leading research on math and code reasoning; 0.76 pass@1 (avg. over 16 samples) with Qwen3-8B. GRPO/PPO advantage shaping; evaluating reasoning by converting chains of thought into graphs and training a GNN.
  • Code agent. Adapted an agentless pipeline for SWE-style tasks and fine-tuned Qwen3-30B-A3B (SFT + RL); owned localization (embedder + LLM) and RL experiments. 0.6 on SWE-bench; also evaluated on internal vLLM repository issues.
  • Data filtration. Adapted NovelSelect to reduce 2M synthetic samples to 20K, improving all target benchmarks (LiveCodeBench, HumanEval).
Yandex — Developer Intern, Alice LLM Assistant (NLP)Jul 2024 – Nov 2024
  • Built a benchmark and training set for the RAG memory model.
  • Trained an embedding model: cos@3 from 0.6 (YandexGPT base embedder) to 0.95; data and approach adopted in production.
  • Addressed LLM hallucinations in the RAG scenario.
HSE Natural Language Laboratory + Yandex — Research InternDec 2023 – present
  • Topological data analysis: approximating first homology with an information-theoretic measure of information storage.
  • The proposed metric proved effective for classifying types of brain activity from EEG data.
HSE University — Research InternApr 2023 – Dec 2023
  • Social-media analysis with mathematical and ML tools: sentiment analysis and topic modeling (BERT, LDA, gradient boosting), discourse analysis with word2vec.
JetBrains Research — Researcher, Cryptography LaboratorySep 2019 – Sep 2020
  • Cryptographic properties of S-boxes; supporting software in C++ and Python (SageMath).

Publications

Education

HSE University, Saint Petersburg2023 – 2025
MSc in Machine Learning and Data Analysis, with Honours · GPA 4.96/5
Thesis: Improved Multimodal Mamba Architecture for Emotion Recognition in Speech. Supervisor: Prof. Ivan P. Yamshchikov. code
Novosibirsk State University2017 – 2021
BSc in Applied Mathematics and Computer Science · GPA 4.6/5
Thesis: Discrete Model of Functioning of a Boolean Network of a Special Structure.

Talks and schools

MNSK talk — dynamics of discrete Boolean networks on the hypercube graph (NSU)Winter 2021
Summer School “Cryptography and Information Security” — one-to-one property and algebraic immunity of S-boxes (NSU)Summer 2020
SEIM’20 talk and paper — S-box construction based on a Boolean function and a permutation (ITMO)Spring 2020
Summer School “Cryptography and Information Security” — construction and analysis of S-boxes (NSU)Summer 2019

Skills

PythonLean 4C++SageMathSFT & RL fine-tuning (GRPO, PPO)Graph neural networks

References

Available on request.