Arman Hatami

PhD Student @ Johns Hopkins University | NeuroAI Researcher
"Breaking down intelligence, one token at a time... or should I say, one neuron?" 🧠
📧 Email 💻 GitHub 🔗 LinkedIn 🎓 Google Scholar

Who Am I? 🤔

I'm a PhD student at Johns Hopkins University working in the NeuroAI field under Prof. Ilya Monosov. I completed my Master's in ECE at JHU (supervised by Prof. Mahyar Fazlyab) and my Bachelor's at Amirkabir University of Technology with a nearly perfect GPA (I was a good student and bad researcher; now I want to be the opposite LOL).

My research focuses on making AI systems more robust, interpretable, and... well, actually intelligent. I work on LLM optimization, unlearning, fine-tuning, and the intersection of NeuroScience and Artificial Intelligence.

🎯 Fun fact: I ranked top 2% among 140,000+ participants in Iran's National Exam. Either I'm pretty good at tests, or 139,200 people had a really bad day.

✨ Currently collaborating closely with the best research partner and life partner: Romina Aalishah ❤️

Research Interests 🔬

The Big Questions 🌌

These questions keep me up at night. If they keep you up too, let's talk.

Question 1: What Happened to the Turing Test?

We said if an agent passes the Turing Test, it's intelligent. Now, in the era of large language models, we've basically achieved that. What happened? Most of us believe LLMs are not intelligent; they're just next-token predictors.

So what is the REAL test? Where can we draw the line and say "this is enough, this agent is actually intelligent"? What's the bar that would convince most of us that a model is trustable at the same level as a human on crucial tasks?

My hypothesis: Perhaps it's the ability to divide and join subtasks. If an agent can break down a problem into the most basic subtasks (where we can't decompose further) AND do the reverse; building up from basics until it reaches complex tasks; maybe THEN we can call it intelligent.

For example: If an agent can explain how to build a bridge starting from F = ma, reasoning through statics, materials science, structural engineering, and finally bridge construction... that's hierarchical reasoning. That's understanding. That's what we're missing.

Question 2: Beyond AGI; What About the Unknown Unknowns?

Everyone's racing toward AGI; a model that performs as well as the best humans in all domains. But that's still a limit. Even if we reach AGI, we're still confined to domains that humans know about.

For Artificial Superintelligence (ASI), we should aim to search in unprecedented domains. How do we know there isn't something like physics or mathematics that we're completely unaware of? What if there's a theory that can explain everything, but we just don't know it exists?

Too many possibilities for our limited human knowledge. So what can we do? How can we make a model that doesn't only rely on known domains? Can we reach an ASI that, based on observations (ours and its own), can explain all events in the universe without any experiments?

🤯 Essentially: Can we build an AI that discovers new fundamental fields of knowledge that humans haven't even conceived of yet?

Let's discuss intelligence, consciousness, ASI, or why my code works on localhost but not in production.

Recent Work 📝

Class Unlearning via Depth-Aware Removal of Forget-Specific Directions
✓ CVPR 2026 – Machine Unlearning in Vision (MUV) Workshop (Oral)
Surgically removing what a vision model knows about a class, layer by layer, without touching everything else it learned.
📄 Paper
Constrained Entropic Unlearning: A Primal-Dual Framework for LLMs
✓ NeurIPS 2025 (Accepted)
Teaching models to forget while staying useful. It's harder than it sounds.
📄 Paper

Internship Experience 💼

I've worked on some pretty cool stuff:

Random Facts 🎲

    ╔══════════════════════════════════════════════════╗
    ║  "I study how models forget.                    ║
    ║   Ironically, I can't forget anything."   ║
    ║                              — Arman            ║
    ╚══════════════════════════════════════════════════╝