Hello, I’m Dmitrii.
I lead fundamental research at White Circle, where we study how AI systems fail and how they behave in out-of-distribution situations. Before that I worked on mechanistic interpretability, as a MATS scholar with Neel Nanda and Arthur Conmy and as a student researcher at ETH Zurich.
Lately I have been thinking about scalable systems for eliciting and evaluating model behavior, the dynamics of agent swarms, and what makes up a model’s perception of its context and task.
Research
Evolves reusable attack strategies against production image generators (DALL·E 3, Nano Banana 2, GPT-Image-2), reaching up to 13% human-verified attack success where prior methods fall to near zero.
Fluid Reasoning Representations
EMNLP 2026 FindingsIn QwQ-32B, representations of obfuscated actions become more abstract as reasoning proceeds; injecting the refined representations early improves accuracy.
Scales sparse feature circuit finding to Gemma-1 2B and finds SAE features that encode which task to perform in context and can induce it zero-shot.
Quriosity: Analyzing Human Questioning Behavior and Causal Inquiry through Curiosity-Driven Queries
IJCNLP-AACL 202513K naturally occurring questions from search engines, human forums and LLM chats; up to 42% of them are causal.
Sparse autoencoders and ITDA on Flux 1: interpretable features of a large diffusion model that can steer image generation.