Large Language Models | Mobina Kashaniyan

How LLM-Driven AutoML Can Improve Cross-Lingual Handwritten OCR

Thu, 02 Jul 2026 00:00:00 +0000

Handwritten text recognition remains a challenging problem in machine learning, especially when models must work across multiple languages and writing systems. Arabic, English, and Persian scripts each introduce different visual structures, character shapes, stroke patterns, and recognition challenges. These differences make cross-lingual handwritten OCR difficult to scale using traditional manually designed models.

In our paper, “LLM-Driven AutoML for Cross-Lingual Handwritten OCR: Closed-Loop Neural Architecture Search with GPT-5, GPT-4o, and Claude Sonnet 4,” we study how large language models can support fully automated model design for handwritten optical character recognition.

The main idea is to use LLMs as AutoML agents. Instead of relying on manual neural architecture design, domain-specific preprocessing, or human model selection, the proposed pipeline allows large language models to generate, evaluate, and refine OCR architectures in a closed-loop neural architecture search process.

The framework uses GPT-5, GPT-4o, and Claude Sonnet 4 to independently propose and improve neural network architectures for handwritten OCR. The approach is evaluated on Arabic, English, and Persian handwritten text recognition, making it relevant to multilingual OCR, low-resource language processing, document intelligence, and automated machine learning.

This work contributes to research on LLM-driven AutoML, cross-lingual OCR, handwritten text recognition, neural architecture search, multilingual AI, and automated model design.

Researchers working on OCR, AutoML, LLM agents, computer vision, Persian handwriting recognition, Arabic handwriting recognition, or low-resource language technologies may find this work useful.

Paper

LLM-Driven AutoML for Cross-Lingual Handwritten OCR: Closed-Loop Neural Architecture Search with GPT-5, GPT-4o, and Claude Sonnet 4
Mobina Kashaniyan, Amirhossein Ghassemi, Nasser Mozayani
Published in 2025 15th International Conference on Computer and Knowledge Engineering (ICCKE), IEEE.

DOI: https://doi.org/10.1109/ICCKE68588.2025.11273810

PDF : https://mobinakashaniyan.github.io/papers/llm-driven-automl-cross-lingual-handwritten-ocr.pdf

IEEE Xplore: https://ieeexplore.ieee.org/abstract/document/11273810

Keywords

LLM-driven AutoML, handwritten OCR, cross-lingual OCR, multilingual OCR, optical character recognition, neural architecture search, GPT-5, GPT-4o, Claude Sonnet 4, LLM agents, Persian OCR, Arabic OCR, English handwriting recognition, document intelligence, low-resource language processing, automated model design, computer vision.

PerfMamba: Performance Analysis and Pruning of Selective State Space Models

Thu, 02 Jul 2026 00:00:00 +0000

Modern sequence models are becoming increasingly powerful, but their computational cost, memory behavior, and deployment efficiency remain important challenges. This is especially relevant for emerging architectures such as Mamba and selective state space models, which have been proposed as efficient alternatives to Transformer-based sequence models.

Our paper, “PerfMamba: Performance Analysis and Pruning of Selective State Space Models,” presents an empirical study of Mamba-1 and Mamba-2. The work analyzes runtime behavior, resource utilization, computation patterns, memory access, I/O characteristics, and scaling properties across sequence lengths from 64 to 16,384 tokens.

The study shows how different components of selective state space models contribute to performance and efficiency. In particular, the paper examines the SSM component, which plays a central role in Mamba-style architectures, and identifies opportunities for optimization through pruning.

Based on the profiling results, PerfMamba explores pruning low-activity states in the SSM component to improve throughput and reduce memory usage while maintaining model accuracy under moderate pruning.

This work contributes to research on Mamba, selective state space models, performance analysis, runtime profiling, model pruning, model compression, efficient AI, high-performance computing, and hardware-aware optimization.

Researchers working on large language models, sequence modeling, AI systems, neural network acceleration, scalable machine learning, and efficient deep learning may find this work useful.

Paper

PerfMamba: Performance Analysis and Pruning of Selective State Space Models
Abdullah Al Asif, Mobina Kashaniyan, Sixing Yu, Juan Pablo Muñoz, Ali Jannesari

Published in: International Symposium on Benchmarking, Measuring and Optimization
Publisher: Springer, Singapore
Year: 2026
Pages: 27–44

Paper page: https://mobinakashaniyan.github.io/publications/perfmamba/

PDF: https://mobinakashaniyan.github.io/papers/perfmamba-performance-analysis-pruning-selective-state-space-models.pdf

Keywords

PerfMamba, Mamba, Mamba-1, Mamba-2, selective state space models, state space models, SSM, sequence modeling, Transformer alternatives, performance analysis, runtime profiling, benchmarking, resource utilization, memory access patterns, I/O characteristics, scaling analysis, model pruning, state pruning, model compression, efficient AI, high-performance computing, hardware-aware optimization, large language models, AI systems, neural network acceleration.

Citation

Al Asif, A., Kashaniyan, M., Yu, S., Muñoz, J. P., & Jannesari, A. (2026). PerfMamba: Performance analysis and pruning of selective state space models. In International Symposium on Benchmarking, Measuring and Optimization (pp. 27–44). Springer, Singapore.

PerfMamba: Performance Analysis and Pruning of Selective State Space Models

Thu, 01 Jan 2026 00:00:00 +0000

This paper contributes to research on PerfMamba, Mamba-1, Mamba-2, selective state space models, performance analysis, model pruning, benchmarking, efficient AI, and high-performance computing by providing a systematic empirical study of the runtime behavior and optimization opportunities of Mamba-style architectures.

Recent sequence modeling research has introduced selective state space models as efficient alternatives to Transformer architectures. However, their real-world performance behavior, memory access patterns, I/O characteristics, resource utilization, and scaling properties require deeper analysis for effective deployment.

PerfMamba profiles Mamba-1 and Mamba-2 across sequence lengths from 64 to 16,384 tokens. The study analyzes computation patterns, memory behavior, I/O characteristics, and scaling trends to identify the components that dominate runtime and resource usage.

Based on these insights, the paper proposes a pruning technique that removes low-activity states within the SSM component. This supports improved throughput and reduced memory usage while maintaining accuracy under moderate pruning.

This work is relevant to researchers working on Mamba, selective state space models, state space models, large language models, sequence modeling, model pruning, model compression, runtime profiling, AI systems, neural network acceleration, and hardware-aware optimization.

Keywords: PerfMamba, Mamba, Mamba-1, Mamba-2, selective state space models, state space models, SSM, sequence modeling, Transformer alternatives, performance analysis, benchmarking, runtime profiling, resource utilization, memory access patterns, I/O characteristics, scaling analysis, model pruning, state pruning, model compression, efficient AI, high-performance computing, hardware-aware optimization, large language models, AI systems.

Citation:

Al Asif, A., Kashaniyan, M., Yu, S., Muñoz, J. P., & Jannesari, A. (2026). PerfMamba: Performance analysis and pruning of selective state space models. In International Symposium on Benchmarking, Measuring and Optimization (pp. 27–44). Springer, Singapore.

LLM-Driven AutoML for Cross-Lingual Handwritten OCR: Closed-Loop Neural Architecture Search with GPT-5, GPT-4o, and Claude Sonnet 4

Tue, 28 Oct 2025 00:00:00 +0000

This paper contributes to research on handwritten OCR, cross-lingual OCR, multilingual AI, AutoML, neural architecture search, and LLM-driven model design by introducing a fully automated pipeline for handwritten optical character recognition across multiple scripts.

Unlike traditional OCR approaches that rely on manual neural architecture design, language-specific preprocessing, or expert-guided model selection, this work uses large language models to generate, evaluate, and refine OCR architectures in a closed-loop neural architecture search process.

The proposed framework applies GPT-5, GPT-4o, and Claude Sonnet 4 to automate model design for Arabic, English, and Persian handwritten text recognition. By testing the method across scripts with different visual and linguistic characteristics, the paper highlights the potential of LLM-driven AutoML for scalable multilingual OCR and low-resource language applications.

This work is relevant to researchers working on optical character recognition, handwritten text recognition, multilingual document intelligence, neural architecture search, AutoML, LLM agents, computer vision, and low-resource language processing.

Keywords: handwritten OCR, cross-lingual OCR, multilingual OCR, optical character recognition, handwritten text recognition, AutoML, neural architecture search, LLM-driven AutoML, large language models, GPT-5, GPT-4o, Claude Sonnet 4, LLM agents, automated model design, document intelligence, multilingual AI, Arabic handwriting recognition, Persian handwriting recognition, low-resource language processing, computer vision for text recognition.

Citation:

Kashaniyan, M., Ghassemi, A., & Mozayani, N. (2025). LLM-driven AutoML for cross-lingual handwritten OCR: Closed-loop neural architecture search with GPT-5, GPT-4o, and Claude Sonnet 4. In 2025 15th International Conference on Computer and Knowledge Engineering (ICCKE), IEEE, pp. 1–6. https://doi.org/10.1109/ICCKE68588.2025.11273810