Research
Work on computer-use agents, autonomous evaluation, vision-language models, Ukrainian NLP, and accessible interfaces.
arXiv preprint
Introduces a noise-corrected reward estimator that enables reinforcement learning from imperfect vision-language evaluator feedback across desktop-agent benchmarks.
ICML 2026 Mechanistic Interpretability Workshop
Studies layer-wise optimal-transport signals for detecting hallucinations in neural machine translation and abstractive summarization.
Research report based on the UNLP 2026 Shared Task
Investigates two-stage retrieval with query-rephrasing and answer-retry loops for an offline Ukrainian retrieval-augmented generation pipeline.
HEAL @ CHI 2026
Evaluates vision-language models as autonomous auditors across macOS, Windows, and Linux benchmarks, covering accuracy, confidence calibration, and inter-model agreement.
TrustAgent @ AAAI 2026
Presents a screenshot-based evaluator for computer-use agents and a dataset of 1,260 human-labeled tasks across 42 built-in macOS applications.
arXiv preprint
Introduces a vision-only framework that reconstructs real-time, tree-structured accessibility metadata from a screenshot.
Proceedings of the Fourth Ukrainian Natural Language Processing Workshop (UNLP 2025)
Introduces a Ukrainian-language benchmark for gender bias in hiring and evaluates prompting, embedding debiasing, and fine-tuning approaches.