Selected work

Representative Papers

GhazalBench: Usage-Grounded Evaluation of LLMs on Persian Ghazals

Feb 06, 2026

While current large language models demonstrate a capacity to comprehend the poetic essence of Persian ghazal poetry, they struggle to reproduce its culturally normative surface form in open-ended generation. This work proposes GhazalBench, a novel evaluation benchmark that, for the first time, incorporates the ability to generate culturally conformant textual forms as a core assessment dimension. The benchmark introduces two tasks: prose-to-poetry comprehension and cue-guided reconstruction of normative verses, complemented by a comparative experiment using English sonnets. Findings reveal that mainstream multilingual models generally excel at semantic understanding but underperform in open-generation of structurally and culturally compliant ghazals. Discriminative tasks notably narrow the performance gap, and the observed limitations are primarily attributed to insufficient coverage of relevant training data rather than inherent architectural constraints.

2 citationsRead paper

Optimizing Urban Critical Green Space Development Using Machine Learning

Jan 01, 2025Sustainable cities and society

Addressing the lack of clear prioritization in urban green space development and the insufficient integration of ecological benefits with social equity—particularly in megacities like Tehran—this study proposes the first quantitative framework for assessing functional importance of urban green spaces by fusing heterogeneous multi-source data (e.g., time-series remote sensing imagery, point-of-interest data, and demographic statistics). Methodologically, it innovatively integrates graph neural networks, multi-objective Bayesian optimization, and remote sensing semantic segmentation to formulate a joint optimization model that jointly maximizes ecosystem services (e.g., urban heat island mitigation and biodiversity support) while enforcing social equity constraints. Empirical validation across five megacities demonstrates a 37% improvement in green space spatial allocation efficiency, a 2.4-fold increase in service coverage for low-income communities, and an F1-score of 91.2%. The framework provides a transferable methodological foundation for sustainable, equity-aware urban green infrastructure planning.

2 citationsRead paper

Multi-modal wound classification using wound image and location by Xception and Gaussian Mixture Recurrent Neural Network (GMRNN)

May 12, 2025

To address insufficient diagnostic accuracy for acute and chronic wounds in clinical practice, this paper proposes a multimodal intelligent classification method integrating wound images and anatomical location information. We introduce a novel collaborative architecture combining Xception for deep visual feature extraction and a Gaussian Mixture Recurrent Neural Network (GMRNN) — the first to explicitly model temporal semantic relationships among anatomical locations. Multimodal features are fused via concatenation and jointly optimized through end-to-end training, overcoming limitations of single-image modality. Evaluated on four wound types—diabetic, pressure, surgical, and venous ulcers—the method achieves classification accuracies ranging from 78.77% to 100%, significantly outperforming conventional deep learning models. This work pioneers the incorporation of anatomical location modeling into intelligent wound diagnosis, empirically validating the efficacy and clinical applicability of location-aware multimodal representation learning.

1 citationsRead paper

Assessing Wildfire Susceptibility in Iran: Leveraging Machine Learning for Geospatial Analysis of Climatic and Anthropogenic Factors

Jan 01, 2025Trees, Forests and People

This study addresses the challenge of wildfire risk assessment in Iran’s arid and semi-arid regions. We developed a multi-source geospatial machine learning model integrating climatic, topographic, land-use, and anthropogenic variables. Methodologically, we employed XGBoost coupled with SHAP-based interpretability analysis, Sentinel-2/Landsat remote sensing classification, and GIS-based spatial interpolation. We propose—novel for Iran—the first nationwide, interpretable wildfire susceptibility classification framework explicitly incorporating both natural and human drivers, alongside a drought-adapted, multi-scale feature engineering paradigm. The model achieves an AUC of 0.92; it identifies 12 high-risk hotspot zones and improves spatial prediction accuracy by 27% over conventional logistic regression. These contributions provide a scientifically robust foundation and methodological reference for national wildfire prevention planning in Iran.

1 citationsRead paper

Experimental Study on Automatically Assembling Custom Catering Packages With a 3-DOF Delta Robot Using Deep Learning Methods

May 14, 20242024 32nd International Conference on Electrical Engineering (ICEE)

This study addresses the need for fully automated sorting and boxing of Persian culinary items using a 3-DOF Delta parallel robot. To overcome the lack of domain-specific visual data, we construct the first Persian food product dataset comprising 1,500 annotated images. We propose a model-free, vision-guided grasping method that eliminates reliance on object CAD models or pose priors: it jointly leverages segmentation masks and principal component analysis to fit oriented bounding rectangles and geometrically solve for dual grasp points. The system integrates YOLOv5 for detection, FastSAM for instance segmentation, mask-driven pose estimation, covariance-based eigenvector computation for grasp point selection, and ROS-enabled real-time motion control. Experiments demonstrate an end-to-end grasping success rate exceeding 80%, enabling closed-loop autonomous operation spanning detection, localization, pose estimation, grasp planning, and execution.

1 citationsRead paper

Recent publications

Latest Papers

PERCEPT: A Corpus for POS Tagging and Analysis of Persian-English Code-Mixing

Aug 10, 2026

This study addresses the absence of a Persian–English code-switched corpus annotated with Universal Dependencies (UD) part-of-speech tags, a gap that has hindered linguistic analysis and the development of syntax-aware NLP models for such mixed-language data. To bridge this gap, we introduce PERCEPT, the first large-scale, multi-platform code-switched corpus comprising 6,800 posts collected from X, Instagram, and Digikala. We further propose a large language model–assisted framework for automatic UD annotation, delivering the first publicly available UD-compliant POS tags for Persian–English code-switched text. Human evaluation confirms high agreement between our automatically generated annotations and gold-standard labels. Linguistic analysis reveals that nouns dominate as the primary category involved in code-switching, the positional distribution of switches remains consistent across platforms, and language-triggering effects are markedly stronger in Digikala.

0 citationsRead paper

Domain-Aware Pruning: Sparsity and Domain Generalization via Regularized Probabilistic Masking

Aug 09, 2026

This work addresses the challenge of simultaneously improving model efficiency and out-of-distribution generalization by proposing Domain-Aware Pruning (DAP), a novel framework that uniquely treats neural network pruning as an intrinsic mechanism for domain generalization. DAP employs differentiable pruning via continuous probability masks and introduces a regularization term to suppress domain-sensitive weights, thereby automatically uncovering domain-invariant sparse subnetworks. Notably, it integrates seamlessly into existing domain generalization pipelines without requiring fine-tuning and is agnostic to the underlying algorithm. Extensive experiments across five benchmarks demonstrate that DAP achieves high sparsity while preserving or even surpassing the out-of-distribution performance of the original dense models, along with enhanced adversarial robustness and interpretability.

0 citationsRead paper

PsychoAgent: An Affect-Sensitive Cognitive Architecture for Conflict-Aware Memory in LLM Agents

Aug 07, 2026

This work addresses a critical limitation in conventional large language model (LLM) agents, which retrieve memories solely based on topical similarity while neglecting the pivotal roles of emotional salience and unresolved conflict in human memory retrieval. To bridge this gap, we propose a novel cognitive architecture that decouples memory into dual factual and emotional channels. A conflict-aware executive controller enables human-like memory recall by first filtering emotionally relevant memories via semantic relatedness, then re-ranking them according to emotional salience, all within a retrieval-augmented generation (RAG) framework. This approach is the first to integrate emotional salience and conflict awareness into LLM memory mechanisms, supporting interpretable, emotion-sensitive retrieval and offline memory reorganization. Evaluated across three conflict-laden scenarios, our method achieves a key memory recall rate of 0.933—significantly outperforming baselines—and demonstrates a 0.22 standard deviation improvement in output quality according to human evaluation, confirming its capacity for sustained emotional effects and dynamic memory adaptation.

0 citationsRead paper

Evaluating the Diagnostic Robustness of Vision-Language Models Under Visual and Textual Perturbations

Aug 05, 2026

Standard accuracy metrics inadequately capture the reliability of vision-language models in medical diagnosis. This work introduces an evidence-preserving perturbation framework—encompassing slice reordering, label position swapping, and lesion removal—applied to a histopathology-validated brain MRI dataset to systematically evaluate the consistency and stability of four model families. The study reveals, for the first time, that these models are highly sensitive to both visual sequence order and textual label arrangement: reversing image sequences flips 48.9% of predictions, shuffling label positions induces 67.8% diagnostic inconsistency, and even after complete lesion removal, 76.1% of outputs remain confidently erroneous. These findings underscore hidden clinical risks masked by high nominal accuracy and advocate for a stability-centered evaluation paradigm to complement conventional accuracy-based assessment.

0 citationsRead paper

UHP Detection: LVLMs have their Unique Hallucination Pattern in the Consistency Space

Aug 04, 2026

This work addresses the challenge of hallucinations in large vision-language models (LVLMs), which often generate responses inconsistent with visual content. Existing black-box detection methods, relying on a single consistency metric, struggle to capture the diverse manifestations of such hallucinations. To overcome this limitation, the authors propose the Uncertainty-based Hallucination Profiling (UHP) framework, which models hallucinations as structured uncertainty patterns defined by image/text perturbations and logical polarity. UHP constructs four complementary consistency groups and leverages intra- and inter-group features to train a lightweight classifier for structured hallucination identification. Experiments demonstrate that UHP achieves state-of-the-art performance, improving AUC-ROC by up to 18.72% and AUC-PR by 20.07% on the AMBER and PhD benchmarks, outperforming both black-box and white-box baselines. Moreover, the identified hallucination patterns exhibit strong cross-dataset generalization.

0 citationsRead paper