1. 
   2026

   Skill Constellations: Tracing the Supply Chain of Agent Skills on GitHub

   Fahd Seddik

   Builds the first dated copy network of 2.2M agent-skill adoptions on GitHub and uses it to rank which repositories to audit before risky skills spread.

2. 
   2026 Under review

   Principled Thoughts for Latent Recursive LLM Systems

   Fahd Seddik, Fatemeh Fard

   REST turns four properties of a valid thought (causality, minimality, separability, stability) into differentiable losses added to cross-entropy for latent single- and multi-agent LLM systems. With the same data and compute, it raises accuracy by up to 7.5 points across 7 benchmarks and adds no parameters at inference.

3. 
   2026 ArabicNLP Arabic Natural Language Processing Conference @ EMNLP 2026 Accepted

   FAAMG at AraSeg Shared Task 2026: Cross-Architecture Ensembling for Punctuation-Free Arabic Segmentation

   Gaser Elmasry, Abdulrahman Elbedewy, Abdelrahman Atef, Fahd Seddik, Mohamed Abdelmoniem

   A weighted ensemble of eight AraBERT, SaT, and CAMeLBERT BiLSTM-CRF models that segments Arabic text with no punctuation or paragraph breaks, fine-tuned on only 174 documents. It reaches 86.7 document-macro F1 on the blind test set and places 4th in AraSeg 2026 Subtask 4.

4. 
   2026 FSE ACM International Conference on the Foundations of Software Engineering

   Panther: Faster and Cheaper Computations with Randomized Numerical Linear Algebra

   Fahd Seddik, Abdulrahman Elbedewy, Gaser Sami, Mohamed Abdelmoniem, Yahia Zakaria

   Randomized numerical linear algebra primitives that drop into existing PyTorch layers – 5x speedup over the dense baseline and 75% parameter reduction at negligible accuracy cost on BERT.

5. 
   2026 Under review

   What Makes a Latent Thought Valid? Defining and Auditing Thought Representations for LLMs

   Fahd Seddik, Fatemeh Fard

   Thought representations have four main properties to be valid.

6. 
   2026 Under review

   ARISE: A Repository-level Graph Representation and Toolset for Agentic Program Repair

   Shahd Seddik, Fahd Seddik, Amirrezza Esmaeili, Mahdieh Sadatbenis, Fatemeh Fard

   Builds a multi-granularity repository graph with statement-level data-flow slicing so a program-repair agent can trace where a variable is defined and used in a single call. Mounted on SWE-agent, it resolves 4.7 more points of SWE-bench Lite issues and nearly doubles line-level fault-localization recall.

7. 
   2026 TOSEM ACM Transactions on Software Engineering and Methodology Under review

   Context-Augmented Code Generation Using Programming Knowledge Graphs

   Shahd Seddik, Fahd Seddik, Iman Saberi, Fatemeh Fard, Minh Hieu Huynh, Patanamon Thongtanunam

   Grounds code generation in a programming knowledge graph so the model retrieves precise structural context (call sites, types, idioms) before generating, instead of relying on lossy in-context recall.

8. 
   2025 EMSE Empirical Software Engineering Under review

   Analysis of AdvFusion: Adapter-based Multilingual Learning for Code LLMs

   Amirreza Esmaeili, Fahd Seddik, Yongyi Ji, Fatemeh Fard, Fuxiang Chen

   An empirical analysis of AdvFusion, an adapter-based multilingual training scheme for code LLMs, with a focus on cross-language transfer and pareto trade-offs.

9. 
   2023 ICECCE International Conference on Electrical, Communication and Computer Engineering

   AGS: Arabic GPT Summarization Corpus

   Abdelrahman Atef, Fahd Seddik, Abdulrahman Elbedewy

   An Arabic abstractive-summarisation dataset built via prompt engineering – the first Arabic abstractive corpus labelled by an LLM. Backed our 1st-place AIC-1 (ICMTC) submission.

3D model:  "3D Origami crane"  by JuanG3D ·  CC BY 4.0