Research engineer¶
English · Русский · AI Interview Atlas › AI roles
AI Engineering · Roles covered: Research engineer
Combines experiments with the software that runs them: architectures, data, training and analysis.
On this page: What the role involves · How it is interviewed (0) · Questions (12) · Preparation
What the role involves¶
Research engineering combines experiments with the software that makes them possible. Anthropic's Research Engineer, Pretraining covers architecture, algorithms, data and optimization, scientific experiments and training infrastructure. It includes independent research projects, so “engineer implements, scientist thinks” is a misleading division. This posting names advanced education; requirements should be checked per team, not generalized to all research-engineering roles.
Sources: Research Engineer, Pretraining, Anthropic, retrieved 2026-09-29
Deliverables and skills: reproducible experiments, training or data code and an analysis of what changed and why. The key distinction from applied product work is investigating or improving model capabilities and training methods, although teams overlap.
How it is interviewed¶
No public interview evidence for these roles is recorded yet. A job posting describes the work; it does not show how candidates are interviewed.
Company pages that cover these roles: Anthropic.
Questions¶
Also practise¶
An editorial selection: relevant to the role, but not reported for it.
- Derive scaled dot-product attention and explain how its scale affects softmax gradients.
- Knowledge · LLM fundamentals · ✍ Answer
- Under a fixed training compute budget, how do Chinchilla-style scaling results change the balance between model size and training tokens?
- Knowledge · LLM fundamentals · ✍ Answer
- Walk through a preference-based RLHF pipeline and explain the roles of the reward model and reference-policy penalty.
- Knowledge · Fine-tuning and post-training · ✍ Answer
- Compare DPO with PPO-based RLHF, including the assumptions behind offline preference learning and reasons to collect fresh rollouts.
- Knowledge · Fine-tuning and post-training · ✍ Answer
- Explain group-relative policy optimisation and the trade-offs of estimating advantages without a learned value model.
- Knowledge · Fine-tuning and post-training · ✍ Answer
- Derive the low-rank weight update used by LoRA and propose an experiment for choosing its rank.
- Knowledge · Fine-tuning and post-training · ✍ Answer
- Choose between prompt changes, retrieval and finetuning for a failing AI feature; include data freshness, latency and total cost.
- Applied scenario · Fine-tuning and post-training · ✍ Answer
- Policy reward keeps rising while humans prefer its outputs less. How would you diagnose and limit reward hacking?
- Applied scenario · Fine-tuning and post-training · ✍ Answer
- A long pretraining run suddenly diverges. How do you investigate data, numerics and optimiser state before resuming from a checkpoint?
- Applied scenario · Fine-tuning and post-training
- Design training for a model that cannot fit on one accelerator; explain how you partition state and keep communication affordable.
- System design · Fine-tuning and post-training · ✍ Answer
- Design an experiment for a suspected emergent capability or bias in a large model, with controls against misleading task framing.
- Applied scenario · Evaluation and observability
- In a short unscripted call, propose testable ideas about a language model's behaviour that need no insider access, only careful observation.
- Knowledge · Evaluation and observability
Preparation¶
Preparation route: strengthen mathematics, ML and LLM fundamentals, then the learning path's advanced branch and post-training where relevant. Practise method selection and LoRA experiments. Reproduce a small result with a baseline, controlled ablation and held-out check. State compute limits and failed hypotheses. The atlas is a starting map; eight weeks of application exercises do not establish frontier research readiness.