What the round looks like
You are shown a cache. It supports one eviction policy, least recently used, and the question is how you would change its design so that other policies fit, LFU being the one named. The candidate spent a long time at the start working out what was being asked: was it a service-level cache, a read-through library? It was neither. It was the refactor itself: given this code, what interface separates the storage from the choice of victim, and what would the LFU version of that interface look like.
The hard part, in the candidate's words, was that LRU and LFU implementations differ a lot, and they could not see how to design the methods and the abstract class to cover both. The interviewer hinted, repeatedly, and they arrived at an answer together. A second candidate, replying later, had the opposite interviewer, one who hardly spoke.
What the refactor is testing
It is an object-oriented design question with code on the table. The move is to notice that the cache needs very little from a policy: to be told when a key was accessed, inserted and deleted, and to be asked which key to drop. Everything specific to LRU or LFU lives behind those calls. Once the contract is written down, the questions follow about its edges: is an update an access, is evict asked before or after the new key is stored, how does the cache stay correct if the policy is wrong.
Our read: whether or not an assistant is in the room, this round rewards the candidate who writes the interface before the classes and can say what each method is for. The code is the easy part once the four methods exist.
What is confirmed, and what is not
- Reported once: the phone-screen round described as an AI round; the given LRU implementation; the refactor toward pluggable eviction policies with LFU named; the interviewer's hints. A reply adds a nearly silent interviewer in another instance.
- Mentioned without content: two AI-enabled interviews in the 2027 new-grad process, in three posts.
- Not known: what the AI in the round actually is (an assistant in the environment, or the candidate's own), the language, how the round is scored, and what the new-grad AI-enabled interviews contain.
How we would prepare
- Write the policy interface before any code. Four methods: accessed, inserted, deleted, choose a victim. Say what the cache does with each, and in what order on a put.
- Implement LRU and LFU behind that interface yourself.LFU with counts, with ties broken by recency, is where the abstraction gets tested; if your interface cannot express it without the cache knowing, it is wrong.
- Keep the old behaviour provable. The same operations through the old cache and the refactored one must agree; say how you would check that.
- Rehearse the discussion, not just the code. The round is a conversation with hints. Think aloud, name the trade-off you are making, and ask whether the approach is viable before you commit to it.
Nyrion's Bloomberg section starts with Cache Policies, built from this report. A working LRU cache is given; the first stage is the refactor into a cache and a policy interface, proven against the old class; the second adds LFU with the tie rule; the third adds FIFO and MRU, a script format and a performance target. The given implementation, the refactor and LFU are what the candidate describes; the exact interface, the later policies and the script runner are ours.
How this round compares with other companies' versions, and what every version grades: the AI coding interview guide.
Frequently asked, briefly answered
Which Bloomberg rounds involve AI?
Two kinds are reported. A software engineer technical phone screen described as an AI round, where an implementation is given and the discussion is how to refactor it (one report, July 2026). And, in the 2027 new-grad process, two interviews candidates call AI enabled, mentioned in three posts from September and October 2026 without any content.
What was the task in the phone screen?
An LRU cache implementation was provided, and the discussion was how to refactor it to be generic enough to support other eviction policies, such as LFU. The candidate first thought it was about a service cache or a read-through library, then realised it was the refactor itself.
Where did the candidate get stuck?
On the abstraction: how to design the methods and the abstract class so that LRU and LFU, whose implementations differ a lot, fit behind one interface. They got there with many hints from an engaged interviewer and did not rate their own performance highly.
Does the interviewer help?
It varies. The July candidate had an interviewer who was engaged throughout and solved it with them. A reply from a later candidate says theirs hardly spoke and answered only when asked whether the approach was viable.
Do Bloomberg's classic rounds continue?
Yes. Reports from the same months describe ordinary coding questions and system design in the phone screen and onsite. This page covers only the rounds described as AI-assisted.
Practice problems tagged for Bloomberg
Browse the catalog and open the Bloomberg section. No login required to look; sign in when you want to solve one.