RepostedByteDance logo

ByteDance · Phone screen

In process

ByteDance Trust & Safety Research Scientist Intern: 1 Experience

Research scientist · InternshipReported Nov 3, 2026
Shared byjanebeingjane

A ByteDance (TikTok) Trust & Safety research scientist intern had two rounds in late 2026: projects, JEV questions and hand-written multi-head attention.

At a glance

  • Process: round 1 (projects, ML knowledge, coding) → round 2 two days later (research talk, a frontier direction, team introduction) (report 01).
  • Asked recently: questions on JEV and where it is used; multi-head attention by hand; present two of your papers or projects; a frontier research direction of your choice (report 01, Nov 2026).
  • Difficulty: rated average; MHA felt fairly standard. Passed round 1; waiting after round 2.
  • Practise first: Implement Multi-Head Attention.

Coverage: One PhD candidate's first two interview rounds for a Trust & Safety research scientist internship, after applying online. The reports' titles say TikTok and their company field says ByteDance, which is where it's listed here. The interviews were in October to December 2026. The first-round report, filed as a technical phone interview, was posted on November 3, 2026, and the candidate's second-round report on November 12, 2026.

What came up

The first round was half projects, then questions about JEV and a hand-written multi-head attention. The candidate calls it fairly standard. The first report doesn't spell out JEV. The second round had no coding: half an hour on the candidate's own research, their view of a frontier research topic, and an introduction to the team's work.

Topic Reports Last seen
Research and project presentations (both rounds) 1 of 1 Nov 2026 (report date)
A frontier research direction of the candidate's choice (round 2) 1 of 1 Nov 2026 (report date)
ML coding: multi-head attention by hand (round 1) 1 of 1 Nov 2026 (report date)
JEV questions (round 1) 1 of 1 Nov 2026 (report date)

Questions we found

Behavioral (BQ)

Question Reports Last seen
Present two of your papers or projects, with questions on why you made certain technical choices (practice) 1 Nov 2026 (report date)
Pick a frontier research direction that interests you and share what you know about it and your own ideas (practice) 1 Nov 2026 (report date)

ML Coding

Question Reports Last seen
Implement Multi-Head Attention, written by hand 1 Nov 2026 (report date)

ML Knowledge

Question Reports Last seen
Questions about JEV, such as where it's used (the report doesn't list them) 1 Nov 2026 (report date)

1 detailed candidate timeline

01: Projects, JEV and multi-head attention, then a research conversation

Interview period: 2026 Q4 · Report dates: November 3 and November 12, 2026 · Role: Research scientist intern, Trust & Safety · Degree listed: PhD · Candidate's rating: Positive, average difficulty (both rounds) · Result: Passed round 1; waiting after round 2

First round

  1. About half an hour on the candidate's projects.
  2. The technical half opened with questions about JEV, such as where it's used.
  3. Coding: write MHA (multi-head attention) by hand, which the candidate found fairly standard. They wrote it in Feishu. You can practice this as Implement Multi-Head Attention. The report gives no tensor shapes or output format, so the practice version's forward pass without learned projections is FastPrep's choice.

The candidate thinks the internship would likely be post-training work. A second round was set up two days later.

Second round

  1. The candidate found this round friendly.
  2. The first half hour was on research and projects: a self-introduction, then two papers or projects of the candidate's choosing (practice). Along the way the interviewer asked a few simple questions, such as why the candidate chose certain technical approaches.
  3. Next, the candidate picked a frontier research direction they found interesting and talked about what they knew and what they thought about it (practice). They chose memory.
  4. The last fifteen minutes were the interviewer's introduction to the team's work, which covers about three directions, mainly multimodal post-training and long-horizon agentic tasks.
  5. On return offers, the candidate heard what they call the usual TikTok answer: it depends on headcount and on how the intern performs.

Result

The candidate felt the second round went well and was waiting for the result, but said they would most likely not take the internship.


Common questions

What coding does ByteDance ask research scientist interns?

Multi-head attention written by hand, in this report (report 01).

What is the ByteDance research intern second round like?

Half an hour on two of your papers or projects, then a frontier research direction you choose, then the interviewer's introduction to the team (report 01).

Keep practicing

Sources: Anywhere useful interview info can be found.

0 total reactions. Your reactions: none.
0 comments

Comments

0

No comments yet — be the first to share what you know.