PhD candidate in generative AI
Hello, I'm a PhD student in computer science at Constructor University, Bremen, where my research is focused on generative and diffusion models for discrete data such as proteins and text. I completed my Bachelor's and Master's degrees in computer science at Lomonosov Moscow State University in 2024. Here, I share my professional notes and personal insights from my ongoing research. My hope is that they will be a valuable resource for other researchers and enthusiasts in the field.
LDLM jointly trains the encoder, diffusion model, and decoder with a simple four-part recipe, achieving stronger text generation at 2–13× faster sampling than prior diffusion LMs.
RLD adapts latent diffusion to antibody design by learning redistributed latent spaces, enabling high-quality conditional generation across organisms and chain types.
This work identifies a critical late-stage sampling failure in Gaussian diffusion on discrete data and shows that self-conditioning and q-sampling restore sample quality across text, code, and proteins.
This work distills continuous diffusion language models into a one-step generator by minimizing MMD in the frozen teacher's representation space, outperforming prior few-step methods on OpenWebText and GSM8K.
This work offers a highly efficient one-step generation process, significantly reducing the computational cost and time required for solving image-to-image translation problems.
Introduced a novel diffusion process that progressively smooths token embeddings based on semantic similarity, combining the benefits of continuous and discrete diffusion.
The method provides guided and efficient remasking for masked language models that drastically enhances quality and speed of text and code generation.
Introduced GeomMotif, a benchmark of 57 protein motif scaffolding tasks, revealing how structural and physicochemical properties affect the performance of protein generative models.
Proposed COSMOS, a text diffusion model operating in a compressed and smooth latent space, enabling up to 8× sequence compression and over 2× faster generation while maintaining competitive text quality.
Developed DiMA, a continuous diffusion framework for protein generation that generalizes across diverse protein language models and supports versatile conditional design tasks, achieving strong generation quality and diversity.
Introduced TEncDM, a text diffusion framework operating on contextual language model representations, achieving superior generation quality over existing non-autoregressive diffusion models across multiple text generation tasks.
CoFlow enables few-step text generation by distilling a flow map inside the latent space of a latent diffusion model while the encoder keeps adapting, outperforming prior flow-map language models on OpenWebText and GSM8K at small step budgets.
AIR refines discrete-diffusion antibody humanization at inference time, steering sequences toward higher humanness while preserving binding-relevant properties without retraining.
2025–2027 (expected)
Advised by Professor Dmitry Vetrov.
Constructor University Bremen, Germany
2022–2024
Moscow State University
2020–2022
Yandex School of Data Analysis, Moscow
2018–2022
Moscow State University
2022–2024
Investigated diffusion probabilistic models for text generation in natural-language and de novo protein design. Trained diffusion language models (up to 7B parameters) using configurations up to 32 GPUs. Results published in four accepted A*-ranked conference papers.
2021–2022
Created a self-supervised detector that flags AI-upscaled videos via contrastive learning.
2021
Built an internal diagnostic tool that scans Alice voice-assistant logs to surface conversational ambiguities and other dialog-routing errors, giving engineers a fast, searchable queue of issues to fix.
2020
Developed a benchmarking toolkit that analyzes and compares full-reference, neural-network video-quality metrics.