Code & data
Code accompanying the lab’s papers. Each repository is listed with the work it belongs to; the full list of papers is on the publications page.
Code
Gaze prompts for interactive correction of medical image segmentation with MedSAM.
Gaze Assistance for Efficient Segmentation Correction of Medical Images, IEEE Access 2025 (Q1)
Cross-modal concept bottleneck models: concepts induced from accompanying text instead of manual annotation.
Cross-Modal Conceptualization in Bottleneck Models, EMNLP 2023 (main track)
Synthetic dataset of primitive shapes with generated descriptions, used to evaluate XCBs.
Cross-Modal Conceptualization in Bottleneck Models, EMNLP 2023 (main track)
Multilingual-Synthetic-Scanpaths
Generative model of reading scanpaths trained jointly on 13 typologically diverse languages.
Multilingual Synthetic Scanpaths, ICML 2026 workshop
Vision–language model acting as the concept bottleneck itself, with a lightweight extraction pipeline.
Simulating Concept Bottlenecks with Vision Language Models, ICLR 2026 workshop
Data
The lab runs its own eye-tracking experiments with practising radiologists — both on public medical imaging corpora and on data from partner clinics. No dataset has been released yet. If you would like to collaborate, write to i.pershin@innopolis.ru.