Anthropic Research Engineer Interview Loop
Anthropic's research engineer interview starts with a recruiter call that tests mission alignment, then a hiring-manager call. Next is a CodeSignal assessment. It's one problem across four levels that get harder, and failing a level blocks the next. For research engineers it also covers NumPy and PyTorch, and the "broken neural net" format. Then comes a take-home project, judged on how you work. The last stage is a four-to-five-hour virtual onsite: ML coding and debugging, distributed training and inference design, research and paper discussion, and culture and values. You can use AI to prepare, but not during the assessment or the interviews.
What decides it
The culture and values interview is a top reason for rejection, and research taste separates candidates at the onsite. Strong coding gets you in the room. It doesn't get you the offer.
Practice by round
Easy coding in both forms the screen uses. One is algorithm basics. The other is the building blocks of a model: cross-entropy from scratch, layer normalization, RMSNorm, GELU, and learning-rate warmup with cosine decay. Coding is a filter here, so clear it quickly.
ML coding at onsite level, like scaled dot-product attention, a causal mask, byte-pair encoding merges, and multiple-choice log-likelihood evaluation. Next to them are the multi-level problems the assessment is built on, such as an in-memory key-value store and a bank account system. Practice the ML questions in debug mode too.
The design conversations Anthropic asks about: batching inference on a single GPU while users wait, and checkpointing a training run that lasts for days. Practice the paper discussion and a past-project walkthrough out loud too.
| Case study | Difficulty |
|---|---|
| Inference Batching, One GPU | Medium |
| Checkpointing Multi-Day Training | Medium |
| Evaluation Harness for Weekly Shipping | Medium |
| Serving Under a p99 Budget | Hard |
| Data vs. Model Parallelism | Hard |
The statistics of evaluating models, like whether a 52% win rate is real and how to estimate pass@k. Then LLM concepts: context windows, regression eval sets and prompt injection. Explain how things work, as well as what they are.
| Question | Difficulty |
|---|---|
| Self-Attention Mechanism | Easy |
| LLM Biases and Mitigation | Medium |
| LLM Feature Cost Control | Medium |
| Context Window Management | Medium |
| Regression Eval Set Design | Medium |
| Is a 52% Win Rate Real? | Medium |
| Transformer Complexity Analysis | Hard |
| Prompt Injection & Guardrails | Hard |
| Estimating pass@k From Samples | Hard |