Research
I study how language models use the documents they are given, how to classify sentiment in Persian, and how deep learning segments medical images. I am applying for PhD positions in AI.
In my latest study, open language models repeated a document that contradicted a fact they knew 49–68% of the time. Their reasoning often stated the true fact first. The signal that something was wrong existed and was thrown away before the answer. I want to find out how to read that signal out, cheaply, and act on it only when it changes the outcome.
When the document is wrongStep-by-step “thinking” cost 10–19× the GPU energy in the same study and improved accuracy without documents, yet made no model more resistant to a wrong one. A follow-up in progress on the 8B and 14B models suggests the amount matters: capping reasoning at 128 tokens reduced switches to the document’s answer by about 9 points, at 24–41% of the energy of unlimited reasoning.
Persian has far less labeled data than English, and reviews are written informally. My thesis and the HVB-Persian preprint combine a fine-tuned transformer with lexicon scores. On about 84,000 reviews the hybrid reaches F1 of 0.88, 0.94 and 0.95 on Digikala, Snappfood and the combined dataset.
HVB-PersianStatus is stated for each. A preprint is not peer-reviewed.
Preprint 2025
SSRN · doi:10.2139/ssrn.5371417
A hybrid sentiment classifier for Persian product reviews that fuses a fine-tuned ParsBERT model with VADER lexicon scores, using weights tuned on validation data. F1 of 0.88, 0.94 and 0.95 on Digikala, Snappfood and the combined dataset of about 84,000 reviews. I developed the method, prepared and merged the datasets, implemented the pipeline and ran the experiments.
In preparation
Manuscript
A systematic survey of deep learning architectures across biomedical imaging modalities, covering supervised, semi-supervised and unsupervised methods, evaluation metrics, clinical relevance and the open problems in 3D architectures and standardised evaluation.
Independent study 2026
Independent research project. Code and all model outputs on GitHub
A reproducible measurement study of what open language models do when a retrieved document contradicts a fact they know, with reasoning switched on and off, and what that reasoning costs in tokens, GPU time and energy.
@misc{jaberi2026conflict,
author = {Soroush Jaberi},
title = {When the Document Is Wrong: Knowledge Conflicts and the Cost of Thinking in Qwen3},
year = {2026},
howpublished = {\url{https://github.com/SoroushJaberi/rag-knowledge-conflict}}
}Improving sentiment classification of Persian text using transformer-based and lexicon-based methods.
Karaj Islamic Azad University · 2023 – present
| Selected coursework | Grade |
|---|---|
| Machine Learning | 20 / 20 |
| Neural Networks & Deep Learning | 20 / 20 |
| Image Processing | 19.5 / 20 |
Karaj Islamic Azad University · 2019 – 2023
| Selected coursework | Grade |
|---|---|
| Artificial Intelligence | 20 / 20 |
| Data Mining | 20 / 20 |
| Algorithm Design | 20 / 20 |
| Computer Graphics | 19.5 / 20 |
Karaj Islamic Azad University · 2024 – 2025
Karaj Islamic Azad University · 2023 – 2025
Karaj Islamic Azad University · 2022 – 2023
A free, open course from Python to AI agents · 2026 – present
If my questions sit near your group's, I would be glad to hear from you. The CV has my full record, and academic references are available on request.
jaberi.soroush@gmail.com