Mahmoud Srewa

Mahmoud Srewa

Third-year Ph.D. Student, Electrical Engineering & Computer Science · UC Irvine

Currently working on aligning LLMs with pluralistic human values, agents that evaluate models automatically, vision-language-action agents that drive like people, and privacy-preserving federated learning.

I'm a third-year Ph.D. student at the University of California, Irvine, advised by Prof. Salma Elmalaki in the Pervasive Autonomy Lab. My research is on AI alignment and trustworthy, human-centered AI. I post-train LLMs with RLHF and preference optimization so they serve diverse, federated populations fairly. I also build agentic systems that evaluate models automatically, and vision-language-action agents that drive and assist the way people do.

In summer 2026 I was an AI/ML Research Intern at Apple AIML, where I built PRISM, an agentic framework that discovers failure modes in production assistant models.

Looking for Summer 2027 internships in LLM post-training, agentic AI, and multimodal / VLA models. Let's talk →

CV Google Scholar LinkedIn GitHub Email
WACV 2027 Apple AIML · 2026 NeurIPS 2026 Top Reviewer 7 publications

Research Interests

Pluralistic & federated LLM alignment

Post-training LLMs with RLHF, PPO, DPO, and GRPO so that one model fairly serves many groups, even when their preference data never leaves them.

Agentic & multi-agent systems

LLM agents that plan, generate, and critique test scenarios to evaluate models automatically, and coding agents trained on synthetic multi-step software tasks.

Vision-language-action models

Retrieval-augmented VLA agents that drive in human styles, and adaptive VLA models for personalized exoskeleton rehabilitation.

Private & robust distributed learning

Federated learning that resists poisoning and inference attacks and spreads privacy risk fairly across clients, using encryption where it counts.

News

  • PaperPersonaDrive was accepted at WACV 2027! 🎉
  • AwardRecognized as a NeurIPS 2026 Top Reviewer (main track).
  • TalkInvited back to Apple's LLM Reading Group to give a talk: "From One Reward to Many: Pluralistic Alignment of LLMs."
  • IndustryWrapped up my internship at Apple AIML. I built PRISM, an agentic framework that automatically discovers failure modes in production assistant models.
  • PaperFinP was accepted at PoPETs 2026.
  • IndustryJoined Apple AIML in Seattle as an AI/ML Research Intern.
  • PreprintPersonaDrive is on arXiv, with a project page and demo videos.
  • PreprintAPPA, our adaptive aggregation method for fair federated RLHF, is on arXiv.
  • PaperOur systematic evaluation of preference aggregation was accepted at the NeurIPS 2025 Workshop on Evaluating the Evolving LLM Lifecycle.
  • AwardNamed a UC Irvine DTEI Scholar for building LLM-based assessment of students' reasoning.
  • PaperPresented PluralLLM at HumanSys '25, co-located with CPS-IoT Week 2025 in Irvine. Code is public.
  • MilestoneStarted my Ph.D. at UC Irvine as a Henry Samueli Endowed Fellow, joining the Pervasive Autonomy Lab.
  • PaperOur poisoning-attack mitigation for privacy-preserving FL appeared at IEEE ICC 2024.
  • MilestoneGraduated with an M.Sc. in Computer Science from the University of Alabama (GPA 4.0).
  • PaperPrivacy-preserving V2V charge sharing appeared at IEEE VTC2023-Spring.

Publications

Newest first · Google Scholar
APPA overview: groups score policy rollouts with local reward models; the aggregation server combines their rewards using a fairness index, EMA history, and dynamic weights before the PPO update.
Preprint · 2026

APPA: Adaptive Preference Pluralistic Alignment for Fair Federated RLHF of LLMs

Mahmoud Srewa, Tianyu Zhao, Salma Elmalaki

In federated RLHF, average aggregation leaves the worst-off groups under-aligned, and min aggregation gives up overall quality. APPA watches a fairness index and, when groups diverge, reweights their rewards by each group's recent alignment history. Under-aligned groups get more weight, and no raw preference data is shared. It improves worst-group alignment by up to 28% over average aggregation on Gemma-2, Llama-3.2, and Qwen3. Hover over the figure, or tap "See it adapt", for a toy animation of the mechanism.

arXiv
FinP framework: clients reduce relative overfitting locally while the server performs adaptive model aggregation using a relative fairness rank, defending against membership and source inference attacks.
PoPETs 2026

FinP: Fairness-in-Privacy in Federated Learning by Addressing Disparities in Privacy Risk

Tianyu Zhao, Mahmoud Srewa, Salma Elmalaki

In federated learning, clients with unusual data face more privacy risk than others. FinP evens out that exposure to source-inference attacks by pairing risk-aware aggregation on the server with client-side regularization that curbs overfitting. It cuts the privacy-risk gap between clients by up to 57% while keeping utility within ±1.75%.

Paper arXiv
Federated RLHF pipeline: training groups score rollouts with local PluralLLM reward models and the aggregation server combines their rewards with adaptive alpha aggregation for PPO.
NeurIPS 2025 Workshop

A Systematic Evaluation of Preference Aggregation in Federated RLHF for Pluralistic Alignment of LLMs

Mahmoud Srewa, Tianyu Zhao, Salma Elmalaki

How should a federated RLHF server combine reward signals from different groups? This paper benchmarks min, max, and average aggregation against an adaptive scheme on opinion Q&A tasks, scoring each on both alignment and fairness. Adaptive aggregation reaches a fairness index of ≈0.99 while keeping overall alignment strong.

OpenReview arXiv
PluralLLM: training groups learn a transformer preference predictor locally and an aggregation server combines their weights with federated averaging; evaluation groups measure alignment.
HumanSys '25 · CPS-IoT Week

PluralLLM: Pluralistic Alignment in LLMs via Federated Learning

Mahmoud Srewa, Tianyu Zhao, Salma Elmalaki

User groups jointly train a preference predictor with federated averaging and never share their data. The predictor then serves as a reward model for aligning LLMs with diverse groups. Compared with centralized training, it converges 46% faster and aligns about 4% better, with comparable group fairness.

Paper Code
System model: smart-meter detection nodes send encrypted local parameters to an aggregation server, with keys issued by a key distribution center.
IEEE ICC 2024

Poisoning Attack Mitigation for Privacy-Preserving Federated Learning-Based Energy Theft Detection

Mahmoud Srewa, Michaela F. Winfree, Mohamed I. Ibrahem, Mahmoud Nabil, Rongxing Lu, Ahmad Alsharif

Federated electricity-theft detection for smart grids. The server finds and drops poisoned model updates by computing similarity directly on encrypted updates, so it never sees them and doesn't need two non-colluding servers. Accuracy holds at 0.919 with 50% of nodes poisoned (0.890 undefended).

Paper
System model: demanding and supplying electric vehicles send encrypted requests and costs to a matching server that runs the Hungarian algorithm; a key distribution center issues keys.
IEEE VTC 2023-Spring

Privacy-Preserving V2V Charge Sharing Coordination using the Hungarian Algorithm

Ahmed Bakr, Mahmoud Srewa, Eyuphan Bulut, Kemal Akkaya, Mizanur Rahman, Ahmad Alsharif

Matches electric vehicles that need charge with vehicles that can share it. The matching server sees only encrypted costs and solves the assignment with the Hungarian algorithm, so no one learns drivers' locations or energy amounts. Privacy adds only about 3% server overhead for 220 users.

Paper

Experience

  1. AI/ML Research Intern · Apple AIML

    Jun 2026 – Sep 2026

    Seattle, WA · on-site

    • Designed and built PRISM, an agentic evaluation framework in which LLM agents plan, generate, and critique test scenarios to find and categorize failure modes in production assistant models. It replaced ad-hoc human evaluation.
    • Reached a 55% failure-discovery rate across 1,514 generated scenarios. A mean pairwise token similarity of 0.25 kept coverage diverse.
    • Built and scaled training and evaluation pipelines on Apple's internal ML infrastructure.

  2. Graduate Research Assistant · Pervasive Autonomy Lab, UC Irvine

    Sep 2024 – Present

    Pervasive Autonomy Lab logoAdvisor: Prof. Salma Elmalaki

    • Coding agents: generating hard, multi-step synthetic software tasks to train LLM agents that plan, call tools, and write and run code.
    • Led PersonaDrive (WACV 2027), a retrieval-augmented VLA driving agent. I built a multimodal retrieval head (SigLIP features plus a learned control encoder, trained with triplet loss on mined hard negatives, with per-style FAISS indexes) and LoRA-tuned InternVL2-1B to read retrieved demonstrations in context.
    • Developing personalized exoskeleton rehabilitation with adaptive VLA models that use multimodal sensor data to tailor therapy in real time.
    • Designed federated RLHF frameworks for pluralistic alignment (APPA, PluralLLM), and co-designed FinP to fix group privacy disparities in FL.

  3. Graduate Research Assistant · University of Alabama

    Jan 2023 – Aug 2024

    Tuscaloosa, AL · Advisor: Dr. Travis Atkison

    • Built a privacy-preserving, poisoning-robust federated learning system for energy-theft detection (IEEE ICC 2024).
    • Designed a privacy-preserving V2V energy-sharing protocol using homomorphic encryption and the Hungarian algorithm (IEEE VTC 2023).
    • Mentored three NSF REU undergraduates through research posters in cybersecurity and ML.

  4. Software Engineer · ProcureDox

    Feb 2020 – Dec 2022

    Toronto, ON · remote

    • Led a team of three building a Spring microservice framework that integrates with eight ERP platforms (NetSuite, QuickBooks Desktop/Online, Microsoft Dynamics NAV, Sage Intacct, …), deployed as Dockerized services.
    • Integrations: built the ERP connectors behind one common interface, handling each platform's own API, authentication flow, and data model, so the core product stayed the same whichever ERP a customer used.
    • Security: designed custom authentication and authorization for the services, and ran security assessments of the platform and its third-party integrations.
    • Built the Spring Boot backend for a property-rental and accreditation platform on AWS: REST APIs, secure third-party integrations, a Cassandra schema, and end-to-end pagination.

Teaching

  1. Teaching Assistant · UC Irvine, EECS

    Mar 2025 – Present

    Irvine, CA

    • Currently teaching EECS 112L Organization of Digital Computers Lab. Previously EECS 125 Introduction to Machine Learning for Engineers and EECS 31L Introduction to Digital Design Lab.
    • DTEI Scholar (2025): built an AI-enhanced thinking-assessment framework that uses multimodal LLMs to grade students' reasoning rather than only their final answers.

  2. Teaching Assistant · Alexandria University

    Jul 2018 – Dec 2022

    Alexandria, Egypt

    • Data Structures, Programming, Algorithms, Machine Learning, Computer Security, and Distributed Systems.

Education

University of California, Irvine

Ph.D., EECS (Computer Engineering)

Sep 2024 – Dec 2028 (expected)
GPA 3.94 / 4.0

Henry Samueli Fellow DTEI Scholar

University of Alabama

M.Sc., Computer Science

Jan 2023 – May 2024
GPA 4.0 / 4.0

Alexandria University

B.Sc., Computer Science

Sep 2014 – Jul 2018
GPA 3.9 / 4.0

Ranked #1 in class

Honors & Service

Honors & Awards

  • NeurIPS 2026 Top Reviewer (main track)2026
  • UC Irvine DTEI Scholar Award2025
  • Henry Samueli Endowed Fellowship2024–25
  • Ranked #1, B.Sc. Computer Science, Alexandria University2018

Reviewing

  • NeurIPS (main track)2026
  • NeurIPS Workshop on Diffusion Language Models (DiffuLM)2026
  • NeurIPS Workshop on Algorithmic Collective Action2025
  • IEEE Internet of Things Journal—

Outside the Lab

Outside research, I'm a husband and a proud father. My wife and daughter are my favorite collaborators, and the only reviewers whose comments I accept without a rebuttal.

  • Hometown: Alexandria, Egypt, whose ancient Library was basically arXiv in 300 BC. Like every proper Alexandrian, I hold that our sea breeze and seafood are superior to Cairo's. This claim does not require peer review.
  • Cooking: I make a serious Alexandrian liver & kidney (kebda w kalawi eskandarani). It's a hometown recipe with no shortcuts.
  • Grilling: every single week. It's the most stable training loop I run.
  • Newport Beach: at least twice a week, for the sunset. The experiments don't stop, though: I keep chatting with Claude Code remotely from the sand so they don't get lonely. 😄
  • Cycling, and football (the one you actually play with your feet).
  • Hiking: of course not. Every PhD homepage says "I enjoy hiking," so I'm just replicating the baseline. Results not yet reproduced.
Mahmoud and his wife Mai smiling on a sunny beach lined with palm trees Mai & meBeach day, experiments running in the background. Mahmoud's wife Mai and daughter Ayla hugging and smiling Mai & AylaThe real senior authors. Mahmoud and his daughter Ayla smiling at a playground Ayla & mePlayground duty, my favorite on-call shift.

Visitors