How to Hire AI Research Engineers and Foundation Model Talent

A Research Engineer builds the systems that make research possible and turns research results into working models: training infrastructure, data pipelines, distributed training, evaluation harnesses, and experiment tooling. Strong engineering ability is non-negotiable.

A Research Scientist owns the research question: what to try, why it should work, and what the result means. Publication history and originality carry more weight here.

Member of Technical Staff (MTS) is a lab convention rather than a distinct discipline. Several frontier labs use one flat title across both research and engineering to avoid an internal hierarchy. When a candidate has MTS on their CV, the title tells you where they worked, not what they did — read the work.

Most teams outside frontier labs need research engineers first. Hiring a pure scientist into a team with no training infrastructure is a well-documented way to lose both the hire and a year.

Pretraining engineers and researchers work on data curation at scale, tokenization, architecture, and large distributed training runs. This pool is genuinely small because very few organizations run large pretraining jobs, and the experience cannot be simulated.

Post-training covers supervised fine-tuning, preference optimization and RLHF-family methods, tool-use and reasoning training, and the evaluation work that decides whether any of it helped. This pool is larger, faster-growing, and where most non-lab companies should be recruiting.

Evaluation is emerging as its own specialty. Engineers who can design honest evals for open-ended behaviour are scarce and disproportionately valuable.

Multimodal, speech, and audio research bring their own pipelines and their own pools. Vision, speech recognition, and audio generation researchers rarely move fluidly between each other's subfields, so treat them as separate searches rather than one requisition.

Compute. If you cannot describe the cluster, the allocation model, and who arbitrates it, the search will stall regardless of comp.

Publication and open-source policy. Say explicitly whether people can publish, at what cadence, and who reviews. Ambiguity is read as 'no'.

Colleagues. Named senior researchers on the team are the highest-conversion element of an outreach message in this pool.

Problem ownership. Whether the hire chooses their research direction, contributes to a directed roadmap, or supports someone else's — say which, honestly.

Compensation is real but comparatively rank-ordered: candidates weigh equity credibility and liquidity expectations more than headline totals, especially when leaving a lab.

Read the work before the interview. For scientists, read one paper closely and ask what they would do differently now. For research engineers, read the code.

Test experiment judgement rather than recall: give an underspecified result and ask what experiment they would run next and what would falsify their hypothesis.

For research engineers, run a real systems evaluation. Distributed training debugging, throughput analysis, and data pipeline design predict on-the-job performance far better than an algorithms screen.

For post-training candidates, ask how they would know a fine-tune actually improved the model, including what they would measure and what regression they would expect to see first.

Assess collaboration explicitly. Frontier-quality research is produced by teams, and a strong solo record without collaborative evidence is a genuine risk at small scale.

Post-training and evaluation searches are moving faster and closing more reliably than pretraining searches, simply because the qualified pool is larger.

Candidates leaving labs consistently ask about compute allocation and publication policy before compensation, and they ask early.

Speech, audio, and multimodal searches behave like separate markets — running them as one requisition slows every one of them down.

The most common mistake we see: a company hires a research scientist to build a product capability that actually needed a strong applied AI engineer, then judges the researcher on shipping velocity.