RepostedCitadel logo

Citadel · Phone screen → Onsite

Not reported

Citadel SWE Intern First-Round Questions: 20 Experiences

Software engineer · InternshipReported Oct 26, 2026
Shared byjanebeingjane

Citadel SWE intern and new grad first rounds from 2025 to 2026: one 45-minute coding interview, often building a stateful system such as a key-value store.

At a glance

  • Process: an OA, then a first round of one or two 45-minute interviews: a few minutes of resume or behavioral talk, then about 30 minutes of coding (12, 13, 17, 19, 20). Passing led to three 45-minute virtual interviews (17, 18, 19, 20).
  • Asked recently: building a stateful class or system in 7 of 20 reports (02, 04, 08, 09, 10, 14, 19); time-indexed lookups and K-way merges twice each; three island problems in one screen (12).
  • Difficulty: unfamiliar starter code cost more time than the algorithm in 03, 08 and 13. Classic tasks ran from Trapping Rain Water (18) to a k-transaction stock DP (11).
  • Practise first: Time-Based Key-Value Store, Merge K Sorted Sequences and Task Dependency Ordering.

Coverage: 20 Citadel SWE first-round reports, March 2025 to October 2026: mostly summer 2026 and 2027 internships, two new grad full-time (01, 04), GQS and Commodities Engineering, and one joint Citadel and Citadel Securities process (13).

What came up

The first round is mostly one coding problem, often one where you build something with state: a store, a file system, a game board. Counts are out of 20.

Topic Reports Last seen
Resume or behavioral questions before coding 10 of 20 Aug 2026 (report date)
Build a stateful class or system 7 of 20 Sep 2026 (report date)
Grids, matrices and boards 3 of 20 Q2 2026
Time-indexed lookup 2 of 20 Feb 2026 (report date)
K-way merge 2 of 20 Nov 2025
Dependencies and recursion 2 of 20 Aug 2026 (report date)
Prefix sums 2 of 20 Q2 2026

A candidate can appear in more than one row. Single-report topics (reachability, LC 857, a math formula, DP, API design) are in the question tables.

Questions we found

Build a class or system

Question Reports Last seen
Time-Based Key-Value Store 2 Feb 2026 (report date)
Reference-Counted Deduplicating File System 1 Sep 2026 (report date)
Limit Order Book Matching Engine 1 Jul 2026 (report date)
2048 Simulation and Board Codec 1 Jan 2026 (report date)
Recursive Task Worker with Memoization 1 Nov 2025
LRU Cache 1 Mar 2025

Algorithms / Data Structures

Question Reports Last seen
Merge K Sorted Sequences 2 Nov 2025
Reachable Transfer Positions 1 Oct 2026 (report date)
Task Dependency Ordering 1 Aug 2026 (report date)
Minimum Cost to Hire K Workers (LC 857), reworded 1 Jul 2026
Search Suggestions with a Prefix Tree 1 Q2 2026
Traverse interior matrix cells in row-major order 1 Q2 2026
Centered k-by-k submatrix sums 1 Q2 2026
Implement a variance-like formula in plain Python 1 Q1 2026
Stock Profit with a Transaction Limit 1 Q4 2025
Number of Islands (LC 200) 1 Dec 2025
Number of Distinct Islands (LC 694) 1 Dec 2025
Number of Islands II (LC 305) 1 Dec 2025
Trapping Rain Water (LC 42) 1 Q1 2025

System Design

Question Reports Last seen
Design a course registration schema and APIs 1 Q2 2025

Behavioral (BQ)

Question Reports Last seen
A technically difficult project or recent problem 2 Aug 2026 (report date)
A feature that helped commercialize a product 1 Aug 2026 (report date)
Why Citadel, and how is it different from Citadel Securities? 1 Dec 2025
A previous internship project 1 Q2 2025
Your most impactful project 1 Apr 2025
A time you had a conflict 1 Apr 2025
Trade-offs in program design 1 Mar 2025

20 detailed first-round experiences

01: GQS: can trades reach every position?

Interview: Q4 2026 · Report date: October 26, 2026 · Role: GQS software engineer, new grad full-time

  1. First technical phone screen. Input: Positions = [(-10, 10), (-20, 20)] and Trades = [(-1, 1), (-5, 5)]. Pick a starting position and return whether the trades can reach every listed position: Reachable Transfer Positions.
  2. Repeat, reverse and start rules were not given.

02: A file system that stores duplicates once

Interview: Q3 2026 · Report date: September 8, 2026 · Team: hedge fund side

  1. Round 1: coding right after introductions, no behavioral questions.
  2. The deduplicating file system started with /a/b/c.txt and /a/b/e.txt, both containing hello. How can they share storage? The candidate proposed hashing so duplicate paths point to one copy, like symbolic links. The interviewer wanted a cheap size check first, hashing only when sizes match.
  3. Starter code: a Python FileSystem with a mock hash, writefile(path, data) and readfile(path). They kept three dictionaries (data by path, duplicate-to-original links, size buckets of (hash, original_path)), but still hashed before checking size.
  4. Deletion follow-up, spoken: a shared pool with reference counts, freed at zero.

03: Summer 2027: order the tasks by their dependencies

Interview: Q3 2026 · Report date: August 15, 2026 · Role: Summer 2027 SWE intern · Result: Rejected (blamed on a different interview)

  1. 15 minutes of introductions and one resume question: a technically challenging experience, and a feature that helped commercialize a product. They used two stories.
  2. Coding: return any valid order of tasks, given a map from task name to its prerequisites. Example: Task1: [Task0], Task3: [Task0, Task1, Task2], Task2: [Task0]; Task0, Task2, Task1, Task3 is valid and Task0, Task1, Task3, Task2 is not.
  3. They explained Kahn's algorithm, then lost time because the input parameter could not change. The interviewer suggested integer in-degree counts. A null-pointer bug remained, and the interviewer chose follow-ups over debugging.
  4. Follow-ups: O(V + E), and how to check every task can finish (compare output length with the count of distinct tasks, including Task0, which appears only as a prerequisite). Hand-typed tests added time pressure.

04: A 45-minute matching engine

Interview: Q3 2026 · Report date: July 26, 2026 · Role: SWE, new grad full-time · Platform: CoderPad

  1. Round 1, 45 minutes: implement an order book matching engine.
  2. API, priority, ties and cancellation rules were not shared.

05: LC 857 with an AI assistant allowed

Interview: July 2026 · Report date: July 29, 2026 · Language: C++ · Platform: Zoom and CoderPad

  1. About 20 minutes on the resume and systems knowledge, then about 20 minutes on a reworded Minimum Cost to Hire K Workers (LC 857). An AI assistant was allowed.
  2. Tracing tests took most of the time. Near the end they used an AI-written solution and explained its output; they called it partly solved. A second first-stage interview was not described.

06: Two first-stage interviews: a trie and matrices

Interview: Q2 2026 · Report date: August 24, 2026 · Result: Did not advance

  1. Session 1: a hard problem they solved recently, then search suggestions with a trie.
  2. Session 2: resume talk ran long. Three linked coding questions were planned: walk a matrix's interior, skipping the border rows and columns; sum each cell's k-by-k neighborhood with 2D prefix sums; and a third that was never shown. They reached task two without time to code it. Window and edge rules were not given.

07: A supplied formula in plain Python

Interview: Q1 2026 · Report date: September 3, 2026 · Role: SWE intern (2027 per the post)

  1. No OA. Two back-to-back 45-minute virtual interviews; the second is in the later-rounds post.
  2. In the first, implement a supplied formula resembling variance in Python, without NumPy, pandas or other libraries. The formula was not reproduced. It felt manageable but easy to get wrong in the details.

08: A timestamped key-value store, straight away

Interview: Q1 2026 · Report date: February 18, 2026 · Result: Rejected

  1. The invite came a little over ten minutes after the OA. No introductions or resume questions.
  2. Implement a timestamped key-value store: set(key, val, timestamp) records a value; get returns the closest value at or before the requested time.
  3. The Python starter used typing and an _Entry(ts: int, val: Any); get oddly also took a val. They proposed a per-key timeline with binary search, but the typing slowed them and they did not finish.

09: 2048 and a board codec

Interview: Q1 2026 · Report date: January 27, 2026 · Role: U.S. SWE intern · Result: Passed

  1. First virtual interview: implement 2048 and encode and decode its board as a long long.
  2. Board size, spawning and bit layout were not recalled. They passed a few days later.

10: Look up a stock price by time

Report date: January 12, 2026 · Result: Rejected

  1. Insert (ticker, price, time) records and query a price by ticker and time, using only Java or Python built-in data structures.
  2. A version of time-indexed lookup; exact-time versus latest-earlier lookup, duplicates and missing history were not specified.

11: Stock profit, from one trade to k

Interview: Q4 2025 · Report date: January 22, 2026 · Result: Passed round 1

  1. First phone screen: best profit from one transaction. They tracked the lowest price in one pass.
  2. The interviewer extended the stock profit task: unlimited trades holding at most one share (holding and not-holding DP states), then at most k trades, dp[i][j][0/1]. They ran examples and covered complexity, space and edge cases.

12: Three island problems in one screen

Interview: Q4 2025 · Report date: December 12, 2025 · Role: Summer 2026 SWE intern · Result: Invited to the virtual onsite about a week later

  1. First technical screen, 45 minutes: about ten minutes on why Citadel and how it differs from Citadel Securities.
  2. Thirty minutes of coding: Number of Islands (LC 200), then Number of Distinct Islands (LC 694); near the end they explained Number of Islands II (LC 305) out loud. The interviewer was pleasant.

13: K-way merge inside a C++ template

Interview: Q4 2025 · Report date: November 19, 2025 · Result: Next-round invite the next morning · Company: joint Citadel and Citadel Securities process until team matching

  1. First technical interview, 45 minutes: about ten minutes of resume, then thirty on merging K sorted arrays.
  2. The C++ template had an abstract interface, type aliases and nested shared_ptr objects. Its ownership API took time; the merge went smoothly. The patient interviewer explained it and let them run slightly over.
  3. They closed with complexity. The three-session second round is in the later-rounds post.

14: An object-oriented worker with memoization

Interview: Q4 2025 · Report date: November 15, 2025 · Role: Summer 2026 SWE intern · Result: Rejected three days later

  1. First virtual interview, any language: an object-oriented task worker using recursion and memoization. The interviewer was a quant developer. The candidate expected LeetCode Medium or Hard problems and made little progress.
  2. In their reconstruction, a Task evaluates its dependencies before compute_fn, with a Number leaf. Part 1 implements Add. Part 2 speeds up the worker and caches Fibonacci: Fib(15) = 610, Fib(100) = 354224848179261915075.
  3. Their reconstruction uses lazy dependencies and a cache keyed by the Fibonacci argument, since caching separately built objects fails if Fib objects expand before evaluation.

15: K-way merge with O(k) extra space

Report date: September 27, 2025 · Role: Summer 2026 SWE intern · Result: Superday invite about a month later; rejected after it

  1. Merge K sorted lists using O(k) extra space, not O(n), and without merging pairs divide-and-conquer style.
  2. No node or iterator API was recalled.

16: Commodities: course registration APIs

Interview: Q2 2025 · Report date: August 22, 2026 · Team: Commodities Engineering · Platform: CoderPad

  1. Before interviewing: a 70-minute OA with two questions drawn at random from four. No prompt was shared.
  2. Round 1, design without code: a course registration database schema and APIs, written in CoderPad with relevant details. They also discussed a previous internship project. Round 2 (shopping and authentication) is in the later-rounds post.

17: Prefix sums and binary search

Interview: Q2 2025 · Report date: April 11, 2025 · Result: Invited on April 14 to three 45-minute virtual interviews

  1. First, behavioral: the most impactful project, and a time of conflict.
  2. Coding: an easy-to-medium prefix-sum and binary-search problem not on LeetCode (input and output not recalled). They knew the approach, fought the details and finished near the deadline, with a supportive interviewer. The OA-to-invite gap was about five days.

18: Trapping Rain Water

Interview: Q1 2025 · Report date: April 18, 2025 · Result: Advanced to three 45-minute virtual interviews, then rejected

  1. A phone interview a few days after the OA: Trapping Rain Water (LC 42). It went smoothly.
  2. The three-session virtual stage followed a few days later.

19: LRU cache and design trade-offs

Interview: Q1 2025 · Report date: March 24, 2025 · Result: Next-stage invite about a week later

  1. Resume talk and general trade-offs in program design, then a standard LRU cache, which they called a familiar template.
  2. The invite to three 45-minute virtual interviews came about a week later.

20: An unnamed Medium with a follow-up

Report date: March 12, 2025

  1. First interview, 45 minutes: about 15 minutes on the resume, with detailed follow-ups on a past internship.
  2. An unnamed LeetCode Medium and a follow-up, then complexity and candidate questions.
  3. Another first-stage interview followed, then an invite to three 45-minute virtual interviews within a week. No prompt was named.

Patterns across the reports

  1. Starter code costs time. Unfamiliar scaffolding slowed 08 (Python typing), 13 (a C++ template) and 03 (a fixed parameter).
  2. Same problems, different candidates. Time-indexed lookup came up in 08 and 10, and K-way merge in 13 and 15.
  3. Three 45-minute interviews next. Every report that described the next stage named three 45-minute virtual sessions (17, 18, 19, 20).

Common questions

What does the Citadel SWE intern first round look like?

One 45-minute interview, sometimes two: 10 to 15 minutes of resume or behavioral talk, then about 30 minutes of coding (12, 13, 20), or straight to code (02, 08).

What coding questions does Citadel ask in the first round?

Often a stateful build: a time-based key-value store (08, 10), a deduplicating file system (02), an order book (04), 2048 (09) or an LRU cache (19). Others were classics like islands (12) and Trapping Rain Water (18).

How fast does Citadel reply after the first round?

From the next morning (13) to about a week (12, 19). Report 17's invite came on April 14, three days after the candidate posted.

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.