Alexey Gavrilov
Machine-learning researcher at ITMO University studying discrete representations, controllable generation, and evidence that keeps model behavior inspectable.
About
Alexey Gavrilov is a machine-learning researcher at ITMO University. His work studies how discrete latent representations shape generated text and code, where quality is lost in compressed generation pipelines, and whether intermediate model states can become useful control points. Current research separates reconstruction loss from latent-generation loss and investigates hierarchical latent codes for partial software regeneration. Across both directions, the aim is to make generative systems easier to diagnose, constrain, and evaluate without treating structural proxies as proof of functional correctness. The publication pages on this site connect concise explanations to full paper text, measured evidence, limitations, and reusable citation records. For questions about the papers or potential research collaboration, contact Alexey by email.
Discrete representations, controllable generation, and inspectable evaluation.
Research themes
Two connected research lines, evaluated through explicit evidence and scope boundaries.
Diagnosing quality loss in compressed generation
Separate codec reconstruction limits from failures introduced during latent generation.
Inspectable control for partial code regeneration
Expose what remains fixed, what may change, and which decoded-output checks support the claim.
Evidence-aware evaluation
Keep measured structural proxies distinct from semantic equivalence and functional correctness.
Selected publications
Browse publication guidesEditorial transparency
Editorial summaries, translations, and research guides may use AI-assisted drafting and are reviewed by the site owner against the cited papers. Publication metadata and full paper text remain source-derived.