Syed Naveed Mahmood

Latest

What's Going On

🎉 Happy news! Our ERUF paper, Representation-Aware Unlearning via Activation Signatures: From Suppression to Entity-Signature Erasure, has been accepted to EMNLP 2026 Findings!

Continued the BlackboxNLP reproducibility task paper on attention-sink circuits and implementation sensitivity.

Began experimental work on new research directions in secure, long-horizon AI agents.

Joined Daffodil International University as Contractual Faculty.

Submitted my representation-aware machine unlearning work to ACL Rolling Review.

Completed my B.Sc. in CSE from BRAC University with a CGPA of 3.95/4.00.

Published GradG-AttnSleep at IEEE ICITEE.

Ongoing research

Current Work

LLM Agent Security

MIRAGE-Persist

Studying how security failures persist across long-running, tool-using AI systems and how to evaluate them reliably.

AI Safety · Interpretability

AgentERUF

Representation-aware diagnostics and interventions for injected or manipulated objectives in AI agents.

Mechanistic Interpretability

Attention-Sink Circuits

Reproducing and stress-testing proposed mechanisms behind attention sinks across implementations and settings.

Efficient Language Models

Activation-Space Model Correction

Testing whether compact activation-level interventions can improve reliability in compressed language models.

Trustworthy AI · Model Security

Mixture-of-Experts Robustness

Investigating internal signals for detecting and mitigating compromised behavior in mixture-of-experts models.

AI for Health · EEG

MIST-Sleep

Building interpretable sleep-staging systems that remain dependable across datasets, recordings, and subjects.

Research outputs

Publications and Projects

01
EMNLP 2026 Findings · 2026

Representation-Aware Unlearning via Activation Signatures

A framework for studying machine unlearning through both model behavior and internal representations.

Paper ↗ Code ↗
02
IEEE ICITEE · 2025

GradG-AttnSleep

Dual-mode visual explanations for attention-based sleep-stage classification.

03
BlackboxNLP Reproducibility Task · Paper in Progress

Same Sink, Different Plumbing

A reproduction and reliability analysis of a proposed GPT-2 attention-sink circuit.

04
Research Software

HyperCite

An AI-assisted workspace for exploring, summarizing, and connecting research literature.

Live project ↗
05
Open-source Project

PaperDebate Agent

A local-first research companion for structured paper analysis and scholarly debate.

GitHub ↗

Academic experience

Teaching

2026 – Present

Contractual Faculty

Daffodil International University

Teaching computer science with an emphasis on conceptual clarity, real-world examples, and accessible technical communication.

Fall 2024 – Fall 2025

Undergraduate Teaching Assistant

BRAC University

Supported Data Structures and Electronic Devices courses through laboratory teaching, grading, assessment, and student mentoring.

Contact

I am open to research discussions and academic collaboration in trustworthy AI, LLM safety, interpretability, and AI for health.