Code-space vs token-space masked diffusion: how to compare them
A stage-consistent comparison protocol for code-space and token-space masked diffusion language models when the discrete codec is lossy.
Alexey Gavrilov — research publications, guides, and notes
Research publications, source-verified bibliographic records, technical guides, and focused evidence notes.
Latest posts
Localized code modification with generative models
How to prevent unnecessary whole-function rewriting while preserving enough freedom for a generative model to make the requested code change.
Constrained code generation for software engineering
A practical distinction between grammar constraints, type constraints, preservation boundaries, and behavior-level acceptance checks for generated code.
AI-assisted refactoring: methods and evidence
How to evaluate recent AI-assisted refactoring methods without confusing a plausible generated patch with verified behavior preservation.
Predictable code generation needs a preservation contract
Why deterministic sampling is not enough, and how observable protected properties and acceptance checks make code-generation behavior testable.
How to tell whether a compressed text generator fails in the codec or the generator
A stage-wise guide to separating reconstruction loss, latent-generation loss, and proxy-metric failure in compressed and discrete-latent text generation.
How can AI edit code without regenerating the entire program?
A technical guide to localized code regeneration, preservation boundaries, edit locality, constrained generation, and hierarchical discrete latent control.
Inspectable Control for Structure-Preserving Software Regeneration
FSE Companion 2026 paper on controllable partial code regeneration. Locking coarse codes raised parse rate from 45.3% to 59.1% while 93.6% of editable positions changed.
Where Quality Breaks in Compressed Short-Text Generation: Staged Bottleneck Localization
A staged diagnostic for compressed short-text generation that separates codec reconstruction from latent-generation loss and compares token-space with code-space MDLM under one shared decoded-text scorer.