MLCB 2026 Oral shared first author

3D-Guided Scalable Flow Matching for Generating Volumetric Tissue Spatial Transcriptomics from Serial Histology

Mohammad Vali Sanian*, Arshia Hemmat*, Amirhossein Vahidi*, Jonas Maaskola, Jimmy Tsz Hang Lee, Stanislaw Makarchuk, Yeliz Demirci, Nana-Jane Chipampe, Muzlifah Haniffa, Omer Bayraktar, Lassi Paavolainen, Mohammad Lotfollahi

HoloTea imputes spot-level gene expression from H&E while looking at the neighbouring sections: morphologically matching spots on adjacent slides are retrieved and fused through a lightweight ControlNet, so the conditioning follows anatomical continuity rather than treating each slice as an island.

Talk and People’s Choice Poster Award, Genomics at Scale 2026.

@article{sanian2025holotea,
  title   = {3D-Guided Scalable Flow Matching for Generating Volumetric Tissue Spatial Transcriptomics from Serial Histology},
  author  = {Sanian, Mohammad Vali and Hemmat, Arshia and Vahidi, Amirhossein and Maaskola, Jonas and Lee, Jimmy Tsz Hang and Makarchuk, Stanislaw and Demirci, Yeliz and Chipampe, Nana-Jane and Haniffa, Muzlifah and Bayraktar, Omer and Paavolainen, Lassi and Lotfollahi, Mohammad},
  journal = {arXiv preprint arXiv:2511.14613},
  year    = {2025}
}

NeurIPS 2026 Main track shared first author

ORCA: Hunting Compositional Failures in Text-to-Image Diffusion

Arshia Hemmat*, Amirhossein Vahidi*, Amitis Shidani, Mohammad Vali Sanian, Hesam Asadollahzadeh, Aryan Yazdan Parast, Mohammad Lotfollahi

Compositional failures in text-to-image diffusion come from misaligned information, not missing information. One auxiliary loss aligns the diffusion latent with a low-rank visual target, steered by the T5 minus CLIP residual.

@inproceedings{hemmat2026orca,
  title     = {ORCA: Hunting Compositional Failures in Text-to-Image Diffusion},
  author    = {Hemmat, Arshia and Vahidi, Amirhossein and Shidani, Amitis and Sanian, Mohammad Vali and Asadollahzadeh, Hesam and Yazdan Parast, Aryan and Lotfollahi, Mohammad},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}

EMNLP 2026 Main conference, oral first author

TDGNet: Hallucination Detection in Diffusion Language Models via Temporal Dynamic Graphs

Arshia Hemmat, Philip Torr, Yongqiang Chen, Junchi Yu

Hallucination in diffusion language models is visible in the denoising trajectory before it reaches the output. TDGNet reads that trajectory as a temporal dynamic graph.

SilkRoadNLP 2026 Workshop at EACL 2026 senior author

Unmasking the Factual-Conceptual Gap in Persian Language Models

Alireza Sakhaeirad, Ali Ma'manpoosh, Arshia Hemmat

DivanBench, 315 hand-curated questions separating memorised Persian cultural facts from reasoning about implicit social norms. Every model tested knows more than it can reason with.

@inproceedings{sakhaeirad-etal-2026-unmasking,
  title     = {Unmasking the Factual-Conceptual Gap in {P}ersian Language Models},
  author    = {Sakhaeirad, Alireza and Ma'manpoosh, Ali and Hemmat, Arshia},
  booktitle = {The Proceedings of the First Workshop on {NLP} and {LLM}s for the {I}ranian Language Family},
  pages     = {1--12},
  year      = {2026},
  address   = {Rabat, Morocco},
  publisher = {Association for Computational Linguistics}
}

EACL 2026 Findings

MEENA (PersianMMMU): Multimodal-Multilingual Educational Exams for N-level Assessment

Omid Ghahroodi, Arshia Hemmat, Marzia Nouri, Seyed M. H. Hosseini, Doratossadat Dastgheib, Mohammad Vali Sanian, Alireza Sahebi, Reihaneh Zohrabi, Mohammad Hossein Rohban, Ehsaneddin Asgari, Mahdieh Soleymani Baghshah

The first large-scale Persian multimodal benchmark: roughly 7.5k Persian and 3k English questions with difficulty metadata and recorded student success rates.

IEEE CSICC 2025 Oral first author

Adaptive Chunking for VideoRAG Pipelines with a Newly Gathered Bilingual Educational Dataset

Arshia Hemmat, Mohammad Hassan Heydari, Kianoosh Vadaei, Melika Shirian, Afsaneh Fatemi

CLIP-SSIM adaptive chunking for video retrieval, with EduViQA-Alpha, a bilingual Persian and English educational VideoQA dataset.

IEEE IoT 2025 Oral first author

Advanced Mutation Testing with Zero and Few-Shot Evaluation Using GPT-4

Arshia Hemmat, Fatame Aghababaei, Mohammadreza Sharbaf, Shekoufeh Kolahdouz-Rahimi

ZeMut generates mutants from structured prompts and outperforms PIT and muJava on several hard cases across twenty classic algorithms.

IEEE ICCKE 2025 first author

Delta-Audit: Explaining What Changes When Models Change

Arshia Hemmat, Afsaneh Fatemi

When a model is retrained or swapped, its accuracy may barely move while its reasons change completely. Delta-Audit explains what actually changed between two versions.

NeurIPS 2025 Workshop VLMs for Real-World Data, spotlight shared first author

From Scenes to Semantics: PersianCLEVR for Bilingual 3D Visual Reasoning

Kianoosh Vadaei, Melika Shirian, Arshia Hemmat, Mohammad Hassan Heydari, Ali Mamanpoosh, Afsaneh Fatemi

A bilingual Persian and English benchmark for compositional 3D visual reasoning, testing whether scene understanding survives a change of language.

Frontiers in Computer Science 2025 first author

Research Directions for Using LLMs in Software Requirement Engineering: A Systematic Review

Arshia Hemmat, Mohammadreza Sharbaf, Shekoufeh Kolahdouz-Rahimi, Kevin Lano, Sobhan Yassipour Tehrani

Where language models genuinely help in eliciting and validating requirements, and where ambiguity in natural-language specifications still defeats them.

NeurIPS 2025 Workshop Embodied World Models for Decision Making first author

ScenePhys: Controllable Physics Videos for World-Model Evaluation

Arshia Hemmat, Emad Aghahosseini, Alireza Nasri, Mohammad Hossein Shaker Ardakani, Amirmasoud Rismanchian, Ali Mamanpoosh, Afsaneh Fatemi

382 controllable simulation clips and 1,146 expert-validated questions covering conceptual, numerical and error-detection reasoning, scored under a strict rubric.

IJCNLP-AACL 2025 Main conference first author

VAGUE-Gate: A Plug-and-Play Local-Privacy Shield for Retrieval-Augmented Generation

Arshia Hemmat, Matin Moqadas, Ali Mamanpoosh, Amirmasoud Rismanchian, Afsaneh Fatemi

A client-side epsilon-local differential privacy gate that rewrites retrieved context before it reaches the model, preserving end-to-end guarantees under post-processing across 32 pipeline variants.

Preprint

PrismSSL: One Interface, Many Modalities; A Single-Interface Library for Multimodal Self-Supervised Learning

Melika Shirian, Kianoosh Vadaei, Kian Majlessi, Audrina Ebrahimi, Arshia Hemmat, Peyman Adibi, Hossein Karshenas

A Python library that runs self-supervised learning across audio, vision, graphs and cross-modal data behind one API. On PyPI under MIT, with HuggingFace integration.

NeurIPS 2024 Datasets and Benchmarks Track first author

Hidden in Plain Sight: Evaluating Abstract Shape Recognition in Vision-Language Models

Arshia Hemmat, Adam Davies, Tom A. Lamb, Jianhao Yuan, Philip Torr, Ashkan Khakzar, Francesco Pinto

IllusionBench hides letters, faces and animals inside ordinary scenes. Human annotators read the shapes almost perfectly; leading vision-language models do not.

@inproceedings{hemmat2024hidden,
  title     = {Hidden in Plain Sight: Evaluating Abstract Shape Recognition in Vision-Language Models},
  author    = {Hemmat, Arshia and Davies, Adam and Lamb, Tom A. and Yuan, Jianhao and Torr, Philip and Khakzar, Ashkan and Pinto, Francesco},
  booktitle = {Advances in Neural Information Processing Systems 37, Datasets and Benchmarks Track},
  year      = {2024}
}

IEEE IKT 2024 Oral first author

Leveraging Retrieval-Augmented Generation for University Knowledge Retrieval

Arshia Hemmat, Kianoosh Vadaei, Mohammad Hassan Heydari, Afsaneh Fatemi

A two-stage Persian retrieval pipeline with in-domain models, released alongside UniversityQuestionBench.

IEEE IKT 2024 Poster

Context Awareness Gate for Retrieval-Augmented Generation

Mohammad Hassan Heydari, Arshia Hemmat, Erfan Naman, Afsaneh Fatemi

A gate that decides whether a query needs retrieval at all, plus a vector-candidates method for scalable semantic search that does not call a language model.

Preprint

CLIP Exhibits Improved Compositional Generalization Through Representation Disentanglement

Reza Abbasi, Amirarshia Hemmat, Mahdi Samiei, Mohammad Hossein Rohban, Mahdieh Soleymani Baghshah

How much of CLIP's compositional generalisation can be traced to disentangled representations rather than data scale.