Myron Sukhanov
MenuClose

Myron Sukhanov

Portrait of Myron Sukhanov
Cambridge, UK

Research Engineer · LLM post-training & evaluation · Scientific ML

Passionate research engineer with experience in LLM post-training and evaluation, scientific ML, and distributed training on HPC.

Currently

  • Visiting Researcher, Applied ML University of Cambridge since Sep 2023
  • First-author paper, manuscript in preparation LLM distillation 2026

Research

First author · manuscript in preparation (2026)

Benchmark Gains Without Reasoning Gains: Filtering, Labelling and Weighting Verified Traces for Distillation

Abstract: We compare methods of leveraging a dataset of chain-of-thought (CoT) traces, where an LLM is prompted to come up with novel and/or groundbreaking heuristics for NP-hard problems. Each of the traces was prompted to end with an executable algorithm that is scored by performance and checked for feasibility by a deterministic verifier. We try three ways of distilling knowledge: gate, grade, and weight. The gate keeps only CoTs that lead to valid algorithms, the grade fine-tunes on all the traces but assigns a quality label to each, and the weight scales the training loss of each trace based on the performance of the algorithm.

We find that the gate is the most promising statistically significant recipe among those tested, scoring a 3.0pp accuracy gain on GSM8K against the unlabelled pool and beating the base by 8.0pp. However, when we investigate further, we find that benchmark improvements over the base come from an increased answer emission rate. When comparing the models' performance on answered-only tests, we find that the base beats all fine-tuned models and beats the gate by 8.6pp. This is a signal that problem-specific fine-tuning of already heavily post-trained 4B–8B-parameter models may get an enhanced emission rate at most.

  1. teacher

    Kimi-Dev-72B

    island-evolutionary search over 4 NP-hard families

  2. deterministic verifier

    13,621 (reasoning, code) traces

    score on 11,731 of them

    pooled dataset

    1. gate

      filter

      (strict / top-tertile)

    2. grade

      quality prefix

      (binary / 4-class / decile / continuous)

    3. weight

      score-shaped loss

      (AWR/CRR-like)

  4. DoRA students

    Qwen3-8B
    Qwen3-4B
    Gemma-3-4B

    controls: base, generic-SFT, unlabelled pool

  5. truncation-honest eval

    in-domain, 3 unseen families

University of Cambridge · manuscript in preparation (2026)

Molecule-Agnostic ML Nanocavity Separator

  • Co-designed, built, and trained a molecule-agnostic ML system for single-molecule SERS spectroscopy data, achieving AUROC 0.97 on human-labelled scans of an unseen molecule
  • Achieved +5.3 dB PSNR compared to the previous-best model within the Physics for Sustainable Chemistry and NanoPhotonics groups
illustrative

King’s College, University of Cambridge · Jun 2024

LLMxLaw Hackathon

  • Developed a solution proposed by Google Cloud: Retrieval Augmented Generation (RAG) based LLM solution to help lawyers find critical points within contract negotiation using Vertex AI API and Streamlit

Experience

  1. University of Cambridge

    Sep 2023 – PresentCambridge, UK

    Visiting Researcher, Applied ML (part-time)

    • Co-designed and built a molecule-agnostic ML system for SERS spectroscopy data
    • Co-developed Scanalyser — a library and scripts for ML model training and inference on an HPC cluster

    Jun 2023 – Sep 2023Cambridge, UK

    Research Intern, Applied ML

    • Reduced the model’s training and inference time by at least 10x via hyperparameter tuning and changing the architecture without performance loss
    • Implemented distributed training on the GPU cluster via a custom training loop leveraging computational-graph optimisations

  2. The Innovation Game

    Aug 2024 – Jun 2025Cambridge, UK

    Computational Scientist (part-time)

    • Researched reinforcement learning (RL) on chain-of-thought (CoT) for self-improving reasoning agents
    • Conducted theoretical AI research for the Neural Network Gradient Descent challenge, which rewards researchers for breakthrough optimisers. See the technical paper
    illustrative

  3. Extreme Light Infrastructure, ELI Beamlines

    Jul 2022 – Aug 2022Prague, Czech Republic

    Research Intern, Data Analysis

    illustrative

Education

  1. University of Sheffield

    Sep 2023 – Jun 2026Sheffield, UK

    BSc Computer Science (Artificial Intelligence), full scholarship

    • Grade: First Class Honours
    • Dissertation: Enhancing Reasoning in Language Models via Verified Algorithmic Chain-of-Thought Distillation
    • Supervisor: Prof. Nikos Aletras

  2. Taras Shevchenko National University of Kyiv

    Sep 2020 – May 2022Kyiv, Ukraine

    BSc Physics (discontinued), grade 93%

Certifications

Awards

  • Mar 2017

    17th nationwide, Ukrainian Physics Olympiad

  • Mar 2018

    35th nationwide, Ukrainian Physics Olympiad

  • Feb 2018

    3rd diploma, Ukrainian Programming Olympiad Regional stage

  • Oct 2018

    2nd diploma, Ukrainian Programming Olympiad City stage

  • 2016–20194 consecutive years

    1st overall in Kherson City, Ukrainian Physics Olympiad City Stage

Contact

Open to research and ML engineering roles, available immediately. Remote welcome.

myron.sukhanov@gmail.com