Yulia Volkova

Yulia Volkova

About meCat

I am an independent AI safety researcher working on AI control evaluations. I hold an MSc in Economics from LSE and have 8 years of experience building insider-threat detection and communication-surveillance systems at Behavox for Tier-1 financial institutions. I witnessed firsthand the integration of AI into the regulatory sector, which inspired me to move into AI safety.

This page highlights my projects and occasionally my cat. My favourite way to spend free time is painting or skiing/hiking in the mountains ☆!

View my CV

Book IconWhat I've been up to

Aug 2026
Panel discussion at GovAI

I was a speaker on a panel discussion about the OpenAI x Hugging Face incident, alongside Marius Hobbhahn (Apollo Research), Hamish Hobbs (CLTR), and Lewis Smith (GDM), moderated by Alan Chan (GovAI). Event page

Jul 2026
Judge for the Secret Loyalties AI safety research hackathon

I was invited to serve as a judge for the Secret Loyalties AI safety research hackathon organized by Apart Research and Formation Research.

Jul 2026
Paper accepted at COLM 2026

Our Cooperative AI paper "Commitment To Cooperation With Self-Negotiated Contracts" (with Tim Wyse, Kaitlin Bustos, and Max Kleiman-Weiner) was accepted at the Conference on Language Modeling.

Jun 2026
MATS project talks at LISA and GDM

We were proud to present our MATS project to audiences at LISA and the GDM London office.

MATS project talk at LISA
May 2026
MATS 9.1 Spotlight Talk

Together with the team, gave a talk on our MATS 9.1 project, opening the London event. Watch on YouTube

MATS 9.1 Spotlight Talk
Mar 2026
FellowCast Podcast at Constellation

Invited to record a podcast episode at Constellation about our work on covert malicious fine-tuning. Listen on Spotify

FellowCast Podcast at Constellation
Jan 2026 - September 2026
MATS

Mentor: Francis Rhys Ward. Tyler Tracy and James Lucassen are acting as advisors.

Project: Control Evaluations for Covert Malicious Fine-Tuning.
Submitted to NeurIPS 2026, preprint coming soon!
Check out the poster!

Nov 2025 - Jan 2026

Mentor: Arun Jose

Research focus: Mechanistic interpretability approach to monitoring chain-of-thought faithfulness.
Check out my project webpage!
Status: gathered feedback from Adam Karvonen, Goodfire team, and Uzay Macar, will be integrating this in the MATS 9.1 extension phase.

Sep - Dec 2025

Mentor: Qiyao Wei (MATS 8.0, Cambridge University)

Research focus: We evaluate how various training pressures affect CoT monitorability in LLMs

Jul 2025 - present
  • Mechanistic Interpretability Project
    PI: Callum McDougall, mentor: Nicky P
    Submitted to ICML [See the paper]
    Framework of Residual Stream Decoders as a method of probing model activations for paragraph-scale and document-scale plans.
  • Cooperative AI Project
    Mentor: Max Kleiman-Weiner
    Accepted at COLM 2026 [See the paper]
    Using the Colored Trails framework to study contract-based improvements for AI agents in game-theoretic setting.
Jul 2025

Completed the ARENA curriculum under David Quarrel. Replicated "Thought Anchors" paper by Bogdan et al. as a capstone project.

ARBOx2 Oxford Bootcamp
Aug 2017 - Sep 2025

Senior Data Scientist, ML Development (Jul 2020 - Sep 2025)

Led the development of Insider Threat ML applications from the Data Science team perspective. These applications analyzed workplace text communications and flagged insider-threat indicators for investigation.

  • Oversaw and took full responsibility for bringing ML applications into production
  • Designed ML applications and maintained end-to-end ownership of their production ML pipelines
  • Managed technical implementation, including developing Java backend and Groovy components to integrate ML applications into the production system
  • Directed and oversaw all ongoing production tasks related to ML applications

Selected technical achievements

  • Ran RoBERTa sequence-length experiments that preserved model performance while increasing training throughput by 2× and reducing compute costs by 50%
  • Developed a synthetic-data generation pipeline that automatically created and uploaded test datasets to AWS S3 through Jenkins, modernizing the QA process
  • Created an automated process that reduced ML application development time from 3 days to 10 minutes
  • Consolidated 4 legacy security-monitoring products into 1 scalable solution, reducing backend maintenance requirements

ML Applications Developer (Aug 2017 - Jun 2020)

Key responsibilities: development of multilingual applications and reports to flag statistical deviations and compliance breaches in client data, designing and building infrastructure to support application creation, QA and maintenance.

Behavox Convention

Book IconOther Research

Other philosophy essays can be seen on the Writing page.

Bouncing Cat