Most major AI labs hire across two research tracks: research scientist and research engineer. Candidates often confuse the two — they appear similar from the outside, both involve ML work, both pay well — but the day-to-day work, the interview rubric, and the career arc are meaningfully different. Picking the wrong track for your background and interests means under-preparing for the loop and under-fitting to the role if you get it.
This piece is a side-by-side comparison and a guide to which track to target.
Research scientist: who does it
Research scientists are the people who design experiments, write papers, and shape the technical direction of the lab. Most have PhDs in ML, related fields (math, physics, computer science theory, computational neuroscience), or have built equivalent research track records through industry research positions.
The day-to-day work:
- Reading papers, formulating hypotheses about new techniques to try.
- Designing experiments, often from a paper-shaped initial framing.
- Writing research code (sometimes well-engineered, often quick-and-dirty for experiment iteration).
- Analyzing experimental results and writing up findings — internal documents and external papers.
- Collaborating with research engineers who help scale and productionize the experiments.
Research engineer: who does it
Research engineers are the people who turn research ideas into runnable, scaled, reliable systems. They build and maintain training pipelines, evaluation harnesses, distributed infrastructure that researchers depend on, and the production paths from a one-off experiment to a deployed model.
Most do not have PhDs. The strongest research engineers come from systems-engineering backgrounds with significant ML exposure — distributed systems engineers who learned ML, ML practitioners who got deep in the engineering, infrastructure engineers who pivoted to AI.
The day-to-day work:
- Building distributed training systems — pipeline parallelism, tensor parallelism, sharding strategies, fault tolerance.
- Building evaluation infrastructure — running thousands of evals across model versions, regression detection, dashboarding.
- Optimizing inference — kernel work, quantization, custom CUDA when necessary.
- Collaborating with research scientists to scale their experiments from a single GPU to thousands.
- Triaging production issues that arise from research-grade code shipped to production paths.
The interview difference
| Round | Research Scientist | Research Engineer |
|---|---|---|
| Paper discussion | Deep, often multiple rounds | Lighter, often one round |
| ML coding | Implement primitives by hand | Implement primitives + optimize them |
| Math fundamentals | Probability, optimization, theory | Less math, more systems |
| Distributed systems | Light or absent | Heavy — training pipelines, fault tolerance |
| Research framing | Open-ended problem; propose investigation | Open-ended problem; propose system to enable investigation |
| Behavioral | Research collaboration history | Engineering ownership stories |
The biggest gap: research scientists need to be able to write a paper-quality justification for a research direction. Research engineers need to be able to architect a system that supports many such directions efficiently.
Career trajectories
Research scientist trajectory
- Junior → senior research scientist → staff research scientist → lead/principal/distinguished researcher
- Promotion is paper-and-impact-driven. Strong publications and citations matter.
- Senior researchers often lead small teams of 2-5 scientists + engineers on a research direction.
- Some transition to research management; most stay IC.
Research engineer trajectory
- Senior research engineer → staff → senior staff/principal
- Promotion is impact-driven, often around scaling and infrastructure that enables research progress.
- Senior research engineers often lead infrastructure projects that 5-50 researchers depend on.
- Transition to engineering management or to platform-engineering tracks is common.
Compensation comparison
Research scientists and research engineers at the same level at the same lab tend to be paid similarly in 2026. There is sometimes a small premium for research scientists at top labs, particularly for scientists with strong publication records, but the gap is smaller than candidates often expect. Both tracks command compensation at the upper end of the tech market.
Which to target
Pick research scientist if:
- You have a PhD in ML, related fields, or equivalent research experience.
- You enjoy the open-ended-research mode of work — formulating hypotheses, designing experiments, writing up results.
- You are comfortable spending months on a single research thread that may not pan out.
- You are excited to publish (or to operate in a culture that values internal publications even when external publication is restricted).
Pick research engineer if:
- You have strong systems-engineering background plus ML exposure.
- You enjoy the build-systems-that-scale mode of work — making research move 10x faster by removing infrastructure bottlenecks.
- You prefer well-defined problems with clear success criteria over open-ended research direction.
- You do not have a PhD but have a track record of shipping technically demanding systems.
The hybrid (and harder) case: applied AI / ML engineer
Many AI labs also hire “ML engineer” or “applied AI engineer” roles that sit between the two. These engineers apply existing models to product problems — fine-tuning, evaluation harnesses for product use cases, data pipelines for production training. The interview loop is closer to research engineer than research scientist, but with more product focus.
If you are unsure between research scientist and research engineer, applied / ML engineer is often the right starting point and converts naturally to research engineer over time.
How to prepare differently
Research scientist preparation:
- Read 10-20 recent papers in your area; build the muscle of explaining their methodology and weaknesses.
- Drill probability, optimization, and linear algebra fundamentals.
- Practice ML coding — implement attention from scratch, sampling routines, training loops.
- Have a research story for your career — what you have published or worked on and how it shaped you.
Research engineer preparation:
- Read recent training-infrastructure papers — Megatron, FSDP, ZeRO, DeepSpeed.
- Build comfort with distributed training concepts: pipeline parallelism, tensor parallelism, gradient checkpointing.
- Drill systems engineering — concurrency, fault tolerance, performance debugging.
- Practice ML coding for infrastructure use cases — implement training loops, eval harnesses, inference optimizations.
Frequently Asked Questions
Can I move from research engineer to research scientist?
Yes, but it usually requires building a research track record over 1-3 years (paper contributions, technical leadership on research projects). The transition happens internally at many labs.
Do I need a PhD for research scientist roles?
Effectively yes at most major labs in 2026. There are exceptions for engineers with exceptional industry research records, but the default path is PhD.
Is research engineer easier to break into?
Yes for engineers without PhDs. The bar is high but the criteria are systems-engineering ones that strong industry engineers can meet without an academic background.
Which track has more demand?
Research engineer roles outnumber research scientist roles at most labs by 2-4x. The labs need many engineers per scientist to scale research effectively.
Which track has more long-term career stability?
Research engineer skills transfer broadly across the AI ecosystem. Research scientist skills are more specialized; if a particular research direction loses funding, scientists may need to pivot harder than engineers do.