Valeriya Zelenkova
Valeriya Zelenkova
CV

Interests

  • AI Interpretability
  • AI Safety
  • Computational Neuroscience
  • Neural Dynamics
  • SciComm

Valeriya Zelenkova

PhD Student at University of Bremen

I'm a neuroscientist moving into AI interpretability and safety. After five years in systems and computational neuroscience — using multi-electrode recordings and population-level analysis to study how information is encoded and communicated between brain areas — I've spent the past year transitioning into interpretability research on large language models, where many of the same questions about neural representation and dynamics reappear. Through a SPAR project I used linear probing and activation patching to trace how an LLM internally represents implicit user attributes across a multi-turn conversation. I'm now looking to build with collaborators in the AI safety community, with a particular interest in bridging neuroscience and interpretability.

Education

  • PhD in Neuroscience, University of Bremen 🇩🇪
  • MSc in Cognitive Science, University of Trento & SISSA 🇮🇹
  • BA in Natural Language Processing, Higher School of Economics

Publications

Projects

  • Latent, Localized, Retrieved On Demand: Tracing Implicit User Attributes Across Multi-Turn Conversation

    A SPAR Spring 2026 project on how an instruction-tuned LLM internally maintains an implicit user attribute across a multi-turn conversation. Combining linear probing, an LLM judge, and activation patching on Llama-3.3-70B-Instruct, we find that hinted attributes are decodable across residual stream throughout the conversation, but only the initial hint positions mediate personalized prediction.

    • Interpretability
    • LLMs
    • User Modeling

Talks