Sviatoslav Chalnev
Bio
Sviatoslav (Slava) Chalnev is an AI researcher based in Sydney, Australia, with a background in mechanistic interpretability and AI safety. He studied at The Australian National University and subsequently pursued independent interpretability research funded by two Long-Term Future Fund stipends totaling $75,000, focused on mechanistic interpretability methods and open-source tooling. He participated in the MATS 6.0 program under Arthur Conmy, resulting in the paper "Improving Steering Vectors by Targeting Sparse Autoencoder Features" (arXiv:2411.02193, 2024), which introduced SAE-Targeted Steering (SAE-TS), a method for constructing steering vectors that target specific sparse autoencoder features while minimizing unintended side effects. He also co-authored "A Single Direction of Truth" (arXiv:2507.23221, 2025), demonstrating that a linear probe on an observer model's residual stream can detect and causally steer contextual hallucinations in language models. More recently, Chalnev co-founded Integuide, an AI startup building tools to capture and disseminate expert technician knowledge, which was part of the Startmate Winter 2025 accelerator cohort.
Links
- Personal Website
- -
- Twitter / X
- -
- LessWrong
- -
Grants
from Long-Term Future Fund
from Long-Term Future Fund
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.
Details
- Last Updated
- Mar 23, 2026, 1:29 AM UTC
- Created
- Mar 20, 2026, 2:58 AM UTC