Collaboration
This page exists so neither of us has to do the vague-email dance. If you want to work with me, read it first. It covers what I work on, how I run projects, what I expect, and what I offer in return. The goal is a quick read: serious people should be able to tell if this is a fit — and if it isn’t, you’ll know that just as quickly.
I mainly work with Georgia Tech undergrads and master’s students, plus the occasional external collaborator. If you’re looking for a PhD advisor, that’s not me. For PhD admissions, reach out to Mark Riedl, Director of the Machine Learning Center, or Sudheer Chava, Alton M. Costley Chair of Finance.
Research Areas
Capability Provenance Attribution and unlearning — where capability comes from.
Persona Vectors Activation spaces, steering, and internal representations.
Social Reasoning Theory of mind, belief tracking, evaluation design.
Safety & Misuse Sycophancy, deception, manipulation — measuring failure modes.
Wargaming Open-ended strategic reasoning under uncertainty.
Negotiation Multi-agent cooperation, coalitions, and incentives.
Finance Domain-specific reasoning where “sounds right” is useless.
My research comes down to one question: what do language models actually learn about people? I chase that across seven areas: capability provenance and training-data attribution, persona vectors and internal representations, social reasoning and theory of mind, safety and misuse of social modeling, wargaming and strategic decision-making, negotiation and cooperation, and finance and financial NLP. Reading lists for every area live on a separate page. Read the list for the area you care about before you reach out.
I lead the research, but I’m not your boss. Those are different roles. I set direction, give context, and push the work forward, but I am not here to micromanage people or force them to work. We are working toward a shared goal, and I want collaborators who act like collaborators: challenge my ideas, ask sharp questions, and then go do the work.
The people who do well with me get stuff done. They show up to meetings. They prepare ahead of time. They take ownership of tasks and get them done quickly instead of waiting around until someone pressures them. They ask clarifying questions when needed, but they do not need constant hand-holding. And when they finish something, they can explain what they did, why they did it that way, and how it turned out.
What I care about is what you get done, not how obediently you take orders. I have had people join projects that were already months underway and, in a fraction of the time, earn second authorship because they showed up, asked good questions, did the work, and got things done. People who are reliable, capable, and do not make excuses can move very fast with me.
What does not work is the opposite: missing meetings, talking more than doing, treating AI tools like they are optional, needing every next step spelled out, not trying to improve, or outsourcing simple tasks that you should have been able to handle yourself. I am not going to drag people across the finish line. The desire to excel has to come from you. Without that fire, this will not work.
Plenty of people want the paper, the CV line, the authorship, the whole idea of being a researcher. Far fewer want to do the actual reps that get you there: the reading, the failed runs, the rewrites you would rather skip. I care about whether you will do the reps.
Most communication happens in Slack. Ask questions early. Do not sit on blockers for days. During paper pushes, expect two meetings a week plus the weekly lab meeting. During quieter periods, once a week is usually enough. Either way, come prepared.
I will give detailed feedback. I will not rewrite your draft for you. Your job is to revise until it is good. The point is not just to get one paper out — it is to get better at doing research and better at writing.
When we are pushing, we move fast. Deadlines are real, and the pace can be intense. When the push is over, we recover. This is not endless grind for its own sake. It is focused sprints with real breaks.
If you contribute meaningfully, you are on the paper. I care about contribution, not seniority games.
I do not like siloed research. People in the group should know what other people are working on and talk to each other directly when the work overlaps. But “talk to each other” does not mean outsourcing basic thinking. Try the obvious things first. Then ask good questions.
Take a task, drive it to done. Do not wait for reminders. When it is finished, be ready to explain what you did, why you did it that way, and how it turned out. I will trust you further the faster you show this.
You do not need to arrive perfect. You do need to close gaps quickly — and three things are non-negotiable.
You need to be comfortable working with AI coding agents — Claude Code, Codex, Cursor, whatever your tool is. This is not optional. We are not here to hand-craft artisanal code. We are here to do good research, and code is a means to that end. If you are unwilling to use agents for coding, or unwilling to learn, we are not a match.
Less than that and projects stall. I would rather work with fewer people who are actually present than more people who show up sporadically.
Show up when you say you will. Deliver what you committed to. Communicate early when something is off track. This matters more to me than raw technical skill.
What I expect you to use day-to-day, what I cover, and how Georgia Tech affiliation factors in.
Python / PyTorch for the core work. GitHub for PRs, branches, and issues. Overleaf / LaTeX for papers. Weights & Biases for experiment tracking. Google Docs / Slides for collaboration. AI coding agents — Claude Code, Codex, Cursor — every day.
Bonus: sprint-based workflows, Kanban boards, and ticket systems. Projects run on sprint cycles with tracked tickets.
You do not pay for compute. GPUs, API credits, and infrastructure are on me.
You do need your own AI assistant accounts — ChatGPT Pro, Claude Pro, or equivalent. I treat that as a cost of doing serious research. I’m working on reimbursement; for now assume it’s part of the job.
Lack of publications does not automatically disqualify you. I am open to promising Georgia Tech undergrads and master’s students. What matters is whether you have read the work, can think and write clearly, and are accountable for what you ship.
Research-for-credit is often the cleanest way to make that real — it creates structure, stakes, and a reason to take the work seriously.
I’m not looking for generic enthusiasm. I’m looking for people who have read the work, can explain what they want to do, and can do it without constant supervision. I don’t need a bunch of helpers. I want young scientists.
Flexible is good. Directionless is not. What I offer in return is real responsibility, real respect, and hard problems worth solving. Do strong work and you can earn authorship, trust, and room to grow fast. I reward competence, not compliance.
Before you reach out: read at least one paper, have a specific reaction to it, know your primary area, and be able to name the exact problem you want to work on and the strongest piece of evidence that you can do it. Keep your answers short — parsimony matters, and more words will not rescue a weak answer.
Read this page. Then decide whether we should talk. Don’t send me a generic paragraph about being passionate about AI — tell me what you read, what you think, and what you want to do.
I read every submission. Strong fits get a 15-minute intro call. The vague ones do not.