Publications

MSP: Modality Self-Play for Training Multimodal LLMs Without Paired Cross-Modal Data
Robert Moseley, Md Kaykobad Reza, Ameya Patil, Edward Ayrapetian, M. Salman Asif
Proposed Modality Self-Play (MSP), a framework that enables zero-shot multimodal composition in multimodal LLMs using proxy textual descriptions and learned bridges, without requiring paired cross-modal non-text training data.
Conference on Neural Information Processing Systems (NeurIPS), 2026

NeurIPS 2026

Tutorial: Foundations and Recent Trends in Robust Multimodal Learning
M. Salman Asif, Md Kaykobad Reza
This tutorial provides a comprehensive overview of robust multimodal learning, covering foundational concepts, recent advancements, and open questions for future research.
IEEE International Conference on Image Processing (ICIP), 2025

Tutorial Webpage IEEE ICIP 2025 Website

Robust Multimodal Learning via Cross-Modal Proxy Tokens
Md Kaykobad Reza, Ameya Patil, Mashhour Solh, M. Salman Asif
Proposed Cross-Modal Proxy Tokens (CMPTs) with low-rank adapters to enhance missing modality robustness of multimodal models, outperforming state-of-the-art baselines across five datasets.
Transactions on Machine Learning Research (TMLR), 2025

TMLR arXiv GitHub Webpage Cite

@article{reza2025robust, title={Robust Multimodal Learning via Cross-Modal Proxy Tokens}, author={Md Kaykobad Reza and Ameya Patil and Mashhour Solh and Salman Asif}, journal={Transactions on Machine Learning Research}, issn={2835-8856}, year={2025}, url={https://openreview.net/forum?id=Wtc6wvcYJ0}, note={} }

MMP: Towards Robust Multi-Modal Learning with Masked Modality Projection
Niki Nezakati, Md Kaykobad Reza, Ameya Patil, Mashhour Solh, M. Salman Asif
Proposed Masked Modality Projection (MMP), which trains a single model to handle missing modalities by learning to project available inputs, improving robustness across various missing-modality scenarios.
IEEE International Conference on Big Data (IEEE BigData), 2025

IEEE Big Data 2025 arXiv Blog Cite

@article{nezakati2024mmp, title={MMP: Towards Robust Multi-Modal Learning with Masked Modality Projection}, author={Nezakati, Niki and Reza, Md Kaykobad and Patil, Ameya and Solh, Mashhour and Asif, M. Salman}, journal={arXiv preprint arXiv:2410.03010}, year={2024} }

Multimodal Learning with Proxy Tokens
Md Kaykobad Reza, Ameya Patil, Mashhour Solh, M. Salman Asif
Proposed proxy tokens combined with low-rank adapters to infer missing modality features from the available ones, thereby improving robustness across diverse missing-modality scenarios.
Asilomar Conference on Signals, Systems, and Computers (Asilomar), 2025

Asilomar 2025 arXiv GitHub Webpage Cite

@article{reza2025robust, title={Robust Multimodal Learning via Cross-Modal Proxy Tokens}, author={Md Kaykobad Reza and Ameya Patil and Mashhour Solh and Salman Asif}, journal={Transactions on Machine Learning Research}, issn={2835-8856}, year={2025}, url={https://openreview.net/forum?id=Wtc6wvcYJ0}, note={} }

Robust Multimodal Learning with Missing Modalities via Parameter-Efficient Adaptation
Md Kaykobad Reza, Ashley Prater-Bennette, M. Salman Asif
Proposed a parameter-efficient adaptation method that enhances missing-modality performance of multimodal models and is applicable to diverse modality combinations, datasets, and tasks.
IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024

IEEE TPAMI arXiv Webpage Blog Cite

@ARTICLE{Reza2024RobustMML, author={Reza, Md Kaykobad and Prater-Bennette, Ashley and Asif, M. Salman}, journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, title={Robust Multimodal Learning With Missing Modalities via Parameter-Efficient Adaptation}, year={2025}, volume={47}, number={2}, pages={742-754}, doi={10.1109/TPAMI.2024.3476487} }

MMSFormer: Multimodal Transformer for Material and Semantic Segmentation
Md Kaykobad Reza, Ashley Prater-Bennette, M. Salman Asif
Proposed a novel fusion strategy and model for multimodal material and semantic segmentation that outperforms state-of-the-art models across all modality combinations.
IEEE Open Journal of Signal Processing (OJSP), 2024

IEEE OJSP arXiv Webpage Code Blog Cite

@ARTICLE{Reza2024MMSFormer, author={Reza, Md Kaykobad and Prater-Bennette, Ashley and Asif, M. Salman}, journal={IEEE Open Journal of Signal Processing}, title={MMSFormer: Multimodal Transformer for Material and Semantic Segmentation}, year={2024}, volume={5}, number={}, pages={599-610}, keywords={Image segmentation;Feature extraction;Transformers;Task analysis;Fuses;Semantic segmentation;Decoding;Multimodal fusion;multimodal image segmentation;material segmentation;semantic segmentation;transformer}, doi={10.1109/OJSP.2024.3389812} }

SSAM: Singular Subspace Alignment for Merging Multimodal Large Language Models
Md Kaykobad Reza, Ameya Patil, Edward Ayrapetian, M. Salman Asif
Proposed a training-free model merging framework, SSAM, that unifies independently trained specialist MLLMs with different input modalities into a single model capable of handling any combination of input modalities.
Under Review

arXiv Cite

@article{reza2026ssam, title={SSAM: Singular Subspace Alignment for Merging Multimodal Large Language Models}, author={Reza, Md Kaykobad and Patil, Ameya and Ayrapetian, Edward and Asif, M Salman}, journal={arXiv preprint arXiv:2603.21584}, year={2026} }

Abstractive Summarization of Bengali Academic Videos Based on Audio Subtitles
Lamisa Bintee Mizan Deya, Farhatun Shama, Abdul Aziz, Md Kaykobad Reza, Md Shahidul Salim
Proposed the first end-to-end pipeline to summarize Bengali academic videos and a new dataset to generate timestamped summaries and titles, enhancing accessibility for a low-resource language.
Findings of EACL 2026

EACL 2026 Cite

@inproceedings{deya2026abstractive, title={Abstractive Summarization of Bengali Academic Videos Based on Audio Subtitles}, author={Deya, Lamisa Bintee Mizan and Shama, Farhatun and Aziz, Abdul and Reza, Md Kaykobad and Salim, Md Shahidul}, booktitle={Findings of the Association for Computational Linguistics: EACL 2026}, pages={6624--6643}, year={2026} }

DualSwinFusionSeg: Multimodal Martian Landslide Segmentation via Dual Swin Transformer with Multi-Scale Fusion and UNet++
Shahriar Kabir, Abdullah Muhammed Amimul Ehsan, Istiak Ahmmed Rifti, Md Kaykobad Reza
Proposed a multimodal model that improves Martian landslide segmentation by separately encoding and effectively fusing RGB and geophysical data, achieving strong performance despite limited labeled samples.
22nd Perception Beyond the Visible Spectrum Workshop (IEEE PBVS) @ CVPR 2026

arXiv Cite

@article{kabir2026dualswinfusionseg, title={DualSwinFusionSeg: Multimodal Martian Landslide Segmentation via Dual Swin Transformer with Multi-Scale Fusion and UNet++}, author={Kabir, Shahriar and Ehsan, Abdullah Muhammed Amimul and Rifti, Istiak Ahmmed and Reza, Md Kaykobad}, journal={arXiv preprint arXiv:2603.14132}, year={2026} }

Model, Analyze, and Comprehend User Interactions within a Social Media Platform
Md Kaykobad Reza, S M Maksudul Alam, Yiran Luo, Youzhe Liu, Md Siam
Developed a graph-based approach to model and analyze social media user interactions, revealing key insights into community dynamics, user behavior, and content preferences.
27th International Conference on Computer and Information Technology (ICCIT), 2024

ICCIT 2024 arXiv Cite

@INPROCEEDINGS{Reza2024Model, author={Reza, Md Kaykobad and Alam, S M Maksudul and Luo, Yiran and Liu, Youzhe and Siam, Md}, booktitle={2024 27th International Conference on Computer and Information Technology (ICCIT)}, title={Model, Analyze, and Comprehend User Interactions within a Social Media Platform}, year={2024}, volume={}, number={}, pages={2104-2109}, doi={10.1109/ICCIT64611.2024.11022085} }

Basin-wide groundwater level forecasting in Kumamoto, Japan: integrating transfer learning with Long Short-term Memory Network
A. T. M. Sakiur Rahman, Shahriar M Sakib, Md Kaykobad Reza, Amiya Basak, Khandaker Nusaiba Hafiz, Md Shazid Islam, Takahiro Hosono
Introduced a basin-wide groundwater level forecasting model that combines transfer learning with LSTM to improve predictive performance across diverse hydrogeological regions.
Hydrological Sciences Journal (HSJ), 2025

Hydrological Sciences Journal Cite

@article{Rahman2025Basin-wide, author = {A T M Sakiur Rahman and Shahriar M. Sakib and Md Kaykobad Reza and Amiya Basak and Khandaker Nusaiba Hafiz and Md Shazid Islam and Takahiro Hosono}, title = {Basin-wide groundwater level forecasting in Kumamoto, Japan: integrating transfer learning with Long Short-term Memory Network}, journal = {Hydrological Sciences Journal}, volume = {0}, number = {ja}, pages = {}, year = {2025}, publisher = {Taylor \& Francis}, doi = {10.1080/02626667.2025.2530120}, URL = {https://doi.org/10.1080/02626667.2025.2530120}, eprint = {https://doi.org/10.1080/02626667.2025.2530120} }

Automatic Summarization of Scientific Articles from Biomedical Domain
Md Kaykobad Reza, Rifat Rubayatul Islam, Sadik Siddique, Md Mostofa Akbar, M Sohel Rahman
Proposed an algorithm that automatically generates summaries and outperforms traditional methods.
International Joint Conference on Computational Intelligence (IJCCI), 2019

IJCCI Paper PDF Slides Cite

@InProceedings{10.1007/978-981-15-3607-6_47, author="Kaykobad Reza, Md. and Islam, Rifat Rubayatul and Siddique, Sadik and Mostofa Akbar, Md. and Rahman, M. Sohel", editor="Uddin, Mohammad Shorif and Bansal, Jagdish Chand", title="Automatic Summarization of Scientific Articles from Biomedical Domain", booktitle="Proceedings of International Joint Conference on Computational Intelligence", year="2020", publisher="Springer Singapore", address="Singapore", pages="591--602", isbn="978-981-15-3607-6" }

Honors & Awards