First time sharing. Calling it an "interview report" is generous — it's more just my interview experience. Background: USNews Top 20 AI PhD, 3rd year, working on agent optimization and agentic RL. My publication record is pretty average, nothing that stands out — 2 ACL + 2 KDD + 1 EACL. I interned at Amazon before this.
03.17 Volcano Ark, AML Enterprise - RS team, HR call. Confirmed I could start any time.
03.20 First round. The first half was chatting about my research, the second half was open-ended questions based on their own systems. They went straight into agent tool use system design — opened by asking what kinds of problems agent tool use runs into in real applications, then broke it down category by category from there: how do you maintain complex state, how do you evaluate a system that complex. Very detailed.
03.25 Second round. They said up front they'd ask 2-3 LeetCode questions, and my legs went weak, had to go change pants. But in the end it was pure research talk the whole time — basically a 90-minute job talk.
03.31 Offer call. The research direction was a good fit and the topic selection seemed pretty flexible, so I accepted.
The biggest personal takeaway (bias, really) is that you need to design and firmly hold onto the pace of the interview yourself. Don't let the topic sit in the interviewer's hands, don't make them do the work of coming up with questions. Build a slide deck that's way more than an hour's worth of material and weave your self-introduction into it. They ask you to introduce yourself? Fine, I'll pull up my slides, and the next slide is my overall research map — well, since we're already on this page, want to hear more about it? Then have a set of specific papers ready, some long, some short, covering different angles — which one are you more interested in? Every step should make it easier for the interviewer to just pick a direction. Let them choose A, B, or C, and then great, now we look at that page of the slide. Do this well enough and unless they have a hard requirement to run through LeetCode or ML fundamentals, they're unlikely to force a stop.
Discussion
Loading comments…