Matey Krastev

Matey Krastev

MSc in Artificial Intelligence at University of Amsterdam

Publications

You can also find my articles on my Google Scholar profile.

Journal Papers

InPars+: Supercharging Synthetic Data Generation for IR

Published in Arxiv pre-print, 2025

This work revisits and extends synthetic query generation pipelines for Neural Information Retrieval (NIR) by leveraging the InPars Toolkit, a reproducible, end-to-end framework for generating training data using large language models (LLMs). We first assess the reproducibility of the original InPars, InPars-V2, and Promptagator pipelines on the SciFact benchmark and validate their effectiveness using open-source reranker and generator models. Building on this foundation, we introduce two key extensions to the pipeline:(1) fine …

Recommended citation: Krastev, M., Hamar, M., Toapanta, D., Brouwers, J., & Lei, Y. (2025). InPars+: Supercharging Synthetic Data Generation for Information Retrieval Systems. arXiv preprint arXiv:2508.13930.
Download Paper

Blog Papers

Towards Enhancing Multi-task Learning for News Recommendation

Published in GitHub, 2024

We reproduce and adapt MTRec, a news recommendation method using pre-trained BERT, for the 2024th edition of the RecSys challenge. MTRec extracts a user representation vector from clicked user articles and scores candidate articles for recommendation by calculating a dot product with each candidate article’s representation. The authors posit auxiliary tasks to aid learning and propose the use of gradient surgery to combine the main task and the auxiliary gradients to the respective losses. In this research, we explore a different auxiliary task, i.e sentiment classification to aid the learning of our task. We further propose to use LoRA instead of full fine-tuning, which we later show to have a regularizing effect and to yield a slightly better performing model than the original authors’ model. Our ablations verify the validity and importance of the included methodological choices.

Recommended citation: Vasilev, S., Krastev, M., & Toapanta, D. (2024). Towards Enhancing Multi-task Learning for News Recommendation.
Download Paper

DynaLoRA: Dynamic Low-Rank Module Allocation

Published in GitHub, 2024

In this project we explored the training dynamics of Parameter-Efficient Fine-Tuning (PEFT) methods, with an emphasis on Low-Rank Adaptation (LoRA). Mainly, we wanted to evaluate, whether it is possible to further reduce memory overhead of fine-tuning by selectively deactivating gradient updates for certain modules during training. In our method, we measured either activation magnitude of the adapted layers in the forward pass, or the gradient magnitude of the same vector in the backward pass.

Recommended citation: Brouwers, J., Fulop, Z., Hamar, M., & Krastev, M. (2024). DynaLoRA: Dynamic Low-Rank Module Allocation.
Download Paper

SegEVOLution: Towards Multimodal Medical Image Segmentation with Context-Prior Learning

Published in GitHub, 2024

This work revisits and extends synthetic query generation pipelines for Neural Information Retrieval (NIR) by leveraging the InPars Toolkit, a reproducible, end-to-end framework for generating training data using large language models (LLMs). We first assess the reproducibility of the original InPars, InPars-V2, and Promptagator pipelines on the SciFact benchmark and validate their effectiveness using open-source reranker and generator models. Building on this foundation, we introduce two key extensions to the pipeline:(1) fine …

Recommended citation: Z. Fülöp, S. Mihailov, M. Krastev, M. Hamar, D.A. Toapanta, S. Achlatis. (2025). SegEVOLution: Towards Multimodal Medical Image Segmentation with Context-Prior Learning.

[RE] Explaining Temporal Graph Models through an Explorer-Navigator Framework

Published in Transactions on Machine Learning 05/2024, 2024

Temporal graphs model complex dynamic relations that change over time, and are being used in a growing number of applications. In recent years, several graph neural networks (GNNs) were proposed, designed specifically for this temporal setting (Xu et al., 2020; Rossi et al., 2020). However, these models are notoriously hard to interpret. For this reason, the original authors (Xia et al., 2023) propose the Temporal GNN Explainer (T-GNNExplainer) – an explorer-navigator framework to efficiently compute sparse explanations of target Temporal GNNs. We reproduce the main findings of the original paper, extend their work by proposing a different type of navigator method, and examine in detail the explanation capabilities and efficiency of the provided framework within various model and hyperparameter settings. We confirm that their explainer outperforms the other baselines across nearly all datasets and metrics. Our findings suggest the navigator helps bias the search process, as well as that T-GNNExplainer can find an exact influential event set. Moreover, we examine the effect of different navigator methods and quantify the runtime-fidelity tradeoff controlled by two hyper-parameters.

Recommended citation: Hamar, M., Krastev, M., Hristov, K. D., & Beglou, D. (2024). [RE] Explaining Temporal Graph Models through an Explorer-Navigator Framework.
Download Paper