Executive Summary
Foundation model organizations have converged on a hiring model that most non-lab employers still misunderstand: flat titles, engineering-heavy research, and evaluation as a first-class discipline.
The practical consequence is that companies competing for this talent with traditional research org structures — publication-weighted expectations, scientist-versus-engineer hierarchy, separate infrastructure teams — lose candidates for structural reasons rather than compensation reasons.
This report is qualitative market intelligence based on Recruits Lab search activity and review of publicly posted research roles at foundation model organizations. It contains no compensation statistics, no headcount statistics, and no survey data, because we hold no dataset supporting them.
Key Findings
- Member of Technical Staff (MTS) is a deliberate flattening. It signals that research and engineering are one job family, and it makes external title-mapping unreliable for sourcing.
- Research Engineer demand exceeds Research Scientist demand in most orgs we see. The bottleneck is people who can run experiments at scale, not people who can propose them.
- Post-training is the tightest sub-market. RLHF-style alignment, preference data, instruction tuning, and eval design require judgment that is hard to acquire outside a lab.
- Pretraining talent is structurally scarce because the experience requires access to large-scale training runs, which few organizations can offer.
- Multimodal, speech, and audio hiring is closer to systems research than to classical ML. Data pipelines, latency, and streaming behavior dominate.
- Publication record is a weakening signal. Evidence of shipped model improvements increasingly outranks venue count.
Market Overview
Three structural facts shape this market.
Compute access is a recruiting instrument. Candidates evaluate offers on what they will be allowed to train and how much they can spend doing it. Organizations that cannot answer this concretely lose late-stage.
Research and production have compressed. Model improvements ship to users on short cycles, so research organizations hire for people comfortable with production consequences. This is why research engineering has become the volume role.
Evaluation has become a career track. Building trustworthy evals for capability, safety, and regression is now specialist work, and it is the area where employers most often have no defined role at all.
Role Definitions
Research Scientist. Sets research direction, formulates hypotheses, designs experiment programs, and owns the intellectual framing of a problem. Typically a PhD or equivalent research track record. Output: novel results and direction.
Research Engineer. Turns research direction into experiments that actually run: training and evaluation infrastructure, data pipelines, distributed training, ablations, reproducibility. Output: reliable experimental throughput and model improvements.
Member of Technical Staff. A single title spanning both, used by several frontier organizations to remove the scientist/engineer hierarchy. Scope is determined by the individual and the team, not by the title.
AI Research Engineer. Common outside the frontier labs; functionally the Research Engineer role, often with more product proximity.
| Dimension | Research Scientist | Research Engineer | Member of Technical Staff |
|---|---|---|---|
| Primary output | Research direction and novel results | Experiments, infrastructure, model improvements | Either, by assignment |
| Typical background | PhD or equivalent research record | Strong systems engineering plus ML depth | Both profiles hired under one title |
| Publication expectation | Higher, though weakening | Low | Varies by org |
| Production expectation | Moderate | High | High |
| Scarcest sub-skill | Problem selection at scale | Large-scale training and eval reliability | Both |
Sub-Market Definitions
Pretraining. Data curation and mixing, tokenization, architecture and scaling decisions, large distributed training runs, and the operational discipline to keep them healthy. Scarce because the experience is gated by compute access.
Post-training. Instruction tuning, preference optimization and alignment techniques, reasoning and tool-use training, distillation, and the eval design that makes any of it measurable. The tightest sub-market we recruit in.
Multimodal research. Vision-language and video-language modelling, cross-modal alignment, and the data engineering underneath it. Heavier on data infrastructure than candidates expect.
Speech and audio AI. ASR, TTS, speech-to-speech, streaming architectures, and latency engineering. Draws from a distinct, smaller community than text-centric research, with strong pockets in academia and voice-native startups.
AI systems research. Inference efficiency, kernels, serving architectures, quantization, and hardware-aware optimization. Overlaps heavily with AI infrastructure hiring.
Hiring Demand: What We Observe
Recruits Lab observations, not measured market data:
Outside the frontier labs, demand is concentrated in post-training and evaluation, because that is where a smaller organization can move a model's usefulness without a pretraining budget. Applied research roles that sit between a model and a product have broadened fastest.
Searches most often stall for two reasons: the role is written as a scientist role but the work is engineering, or the organization cannot describe its compute environment. Both are fixable before going to market.
Skills and Requirements
- Experimental throughput. Ability to design and run many clean experiments, with reproducibility discipline.
- Distributed training fluency. Parallelism strategies, failure recovery, throughput debugging.
- Data judgment. Curation, filtering, contamination awareness, and mixture decisions — repeatedly cited by candidates as where results are actually won.
- Evaluation design. Building measures that survive contact with real usage, and detecting regressions.
- Post-training technique depth where relevant: preference data pipelines, reward modelling, reasoning and tool-use training.
- Inference awareness. Understanding how training choices land in serving cost and latency.
- Written research communication. Internal reports carry more weight than papers in most organizations we work with.
Talent Supply
Sourcing pools, qualitatively:
Frontier and near-frontier labs hold most of the pretraining experience, and movement between them is the primary way that experience circulates. Academic groups remain the deepest source for multimodal, speech, and audio research, and for evaluation methodology. Infrastructure and high-performance computing engineers are an underused source for research engineering, particularly for systems research and inference work. Strong applied ML engineers convert well into post-training roles when the organization can offer mentorship from someone who has run the technique before.
Compensation
This report states no research compensation figures. Frontier research compensation is dominated by equity structures that public aggregators represent poorly, and reproducing those figures as research would be misleading.
For planning, see guides that disclose their basis:
Practitioner observation: at this level, candidate decisions are driven by compute access, colleagues, publication or disclosure policy, and problem ownership at least as much as by cash. Employers that compete only on cash tend to lose to organizations offering scale of work.
Geography
Observation, not measurement: the San Francisco Bay Area holds the deepest concentration, followed by New York, Seattle, London, Zurich, Paris, and Toronto. Speech and audio research is more geographically distributed, with visible European and Asian academic pipelines. Fully remote research hiring is less common than in AI engineering, because compute environments and collaboration density favor co-location.
Experience Requirements
For research engineering, three to eight years of strong systems engineering with demonstrated ML depth is more predictive than a PhD. For research scientist roles, a PhD or an equivalent record of independent results remains the norm. For pretraining specifically, prior involvement in large training runs is close to non-substitutable; teams without it should hire post-training and evaluation first.
Employer Implications
- Decide whether you are hiring direction-setting or experimental throughput, and title it honestly.
- Describe your compute environment concretely in the process. Vagueness reads as weakness.
- State your publication and disclosure policy early; it is a real decision factor.
- Create an explicit evaluation role rather than distributing it across the team by default.
- If you cannot pretrain, compete in post-training, evaluation, multimodal, or speech, where entry is possible.
Candidate Implications
Candidates who document what they trained, what improved, and how they measured it are evaluated faster than candidates who lead with publication lists. For those moving from academia, evidence of production consequence — a model change that shipped — is the strongest differentiator.
Methodology
- Type: Qualitative market intelligence. Practitioner observations, not statistical findings.
- Basis: Recruits Lab search and scoping activity across AI research and research engineering roles, plus review of publicly posted Research Engineer, Research Scientist, and Member of Technical Staff job descriptions at foundation model organizations.
- Timeframe: Reflects searches and postings current as of Q3 2026.
- Sample size: Not applicable. No statistical claims are made.
- Definitions: Ours, synthesized from public postings and hiring-team scoping conversations.
Full framework: how Recruits Lab research is produced.
Limitations
- No first-party quantitative dataset underlies this report; nothing here is a measured market statistic.
- Frontier lab hiring is opaque. Public postings understate internal mobility and referral hiring.
- Observations skew toward organizations that engage external recruiting partners.
- No compensation, headcount, or geographic distribution figures are stated, because we cannot source them credibly.
Sources
- Publicly posted Research Engineer, Research Scientist, and Member of Technical Staff job descriptions at foundation model organizations (reviewed 2026).
- Recruits Lab first-party search and scoping activity (qualitative).
- Recruits Lab salary guides, which state their own sources and limitations.
Related Research
- 2026 Agentic AI Hiring Report
- 2026 Forward Deployed Engineer Talent Report
- 2026 AI Engineer Hiring Report
Related Hiring Guides
- How to Hire AI Research Engineers
- AI Engineer vs ML Engineer vs Research Engineer
- How to Interview AI Engineers