What I Learned From Going Through Meta's AI Researcher Interview Loop

I made it to Meta's final round for an AI Researcher role. Two years of industry research experience, relevant publications, solid ML fundamentals. I thought I was ready.
I wasn't — not because the technical bar was higher than expected, but because I prepared for the wrong version of the interview.
Here's what the loop actually looks like from the inside, and where the real difficulty sits.

The loop structure (briefly)
For context, the Meta AI Researcher loop typically covers:

Coding rounds (ML-adjacent implementation, not pure LeetCode)
Research presentation (deep dive into your own work)
Research discussion (your opinions on the field)
ML system design
Behavioural

Most write-ups focus on the coding. That's not where candidates lose this interview.

Research taste is the hardest thing to prepare for
The research discussion rounds aren't testing whether you've read the papers. They're testing whether you have opinions about the field — independent, defensible views on what matters, what's overhyped, and what's genuinely unsolved.
A question like "what do you think the biggest open problem in X is right now?" is not asking for a literature summary. If your answer sounds like an abstract, the energy in the room shifts immediately.
The researchers interviewing you work on this stuff daily. They want to know if you have something interesting to say about it — even if they disagree with you. Especially if they disagree with you.
Prep for this by actually forming opinions. Read workshop discussions and position papers, not just polished proceedings. Follow the live debates on arXiv and Twitter/X. Know what you find compelling and what you find unconvincing — in your own words.

The research presentation is a defence, not a showcase
Most candidates walk into the presentation trying to impress. That's the wrong frame.
Meta interviewers are reading your papers before the call. They're not there to learn what you built. They want to understand how you think — specifically whether your contributions came from sound scientific reasoning or from fortunate circumstances.
The questions that reveal this most clearly:

"What would you do differently if you started this project today?"
"What's the biggest limitation of this approach?"
"Why did you use X instead of Y here?"

Defensive or vague answers to these are the fastest way to lose the loop. Specific, honest, intellectually curious answers are what they're looking for. Prepare the hardest version of each question. Practise answering them out loud without getting defensive.

Don't underestimate the coding rounds
Research candidates often skip serious coding prep. Big mistake.
The questions aren't pure algorithm problems, but they're not easy either. Expect things like:

Implementing attention or a custom layer from scratch
Debugging a training loop with a subtle numerical issue
Writing clean, efficient data processing code under time pressure

The implicit question throughout is: can you be a self-sufficient researcher in a real codebase? Can you implement your own ideas cleanly? Can you spot why an experiment is behaving unexpectedly?
Practise implementing core ML components from first principles. Get comfortable with PyTorch internals. Don't leave this until the last week.

ML system design at Meta scale is its own thing
This round trips people up because it requires thinking that sits between research and engineering — and most prep resources cover neither well at this level.
Designing a ranking or recommendation system at Meta means thinking about:

Latency budgets: what model complexity can you afford at p99 < 10ms?
Feedback loops: your model's outputs affect future training data. How do you prevent degradation or unintended amplification?
Continuous training: static models go stale fast. How do you handle distribution shift and safe rollout?
Integrity and fairness: for many Meta product areas, the ML system design has direct implications for what content gets surfaced. Strong candidates raise this unprompted.

Study Meta AI's engineering blog posts and industrial ML papers alongside the system design prep. Practice designing systems end-to-end, out loud, with someone who will ask the operational questions you haven't thought through.

The honest summary
The Meta AI Researcher interview is genuinely hard in the right ways — it's testing the things that actually make a good researcher, not just proxies for them. The candidates who do well aren't necessarily the ones with the most impressive CVs. They're the ones who have developed real opinions, can defend their work honestly, and think clearly about systems at scale.
I'd approach the prep as four separate workstreams — research presentation, research opinions, coding, and system design — rather than one generic interview prep track. Each requires different preparation, and the overlap is smaller than it looks.
Happy to answer questions in the comments if you're actively prepping for this.

Comments

[klement Gunndu]

The distinction between research taste and research knowledge is sharp. Do you think the emphasis on forming independent opinions would work against candidates who genuinely prefer collaborative research over having strong individual takes?