LLM Research Scientist (Pre-training & Computer Vision & Adversarial Robustness)

$100–120/hrrate as listed

Mercor is staffing ML researchers to run scoped empirical experiments across vision and language — training image classifiers and generative models from scratch, fine-tuning open-weight LLMs, and hardening models against adversarial inputs and adversarial conversations. Work is project-based and remote, with observed rates of $100–120/hr depending on depth in a specialty area.

Summary, eligibility guidance and screening preparation are prepared by Sidequest from the platform's posting. Check the source listing before applying.

What the work involves

These are well-scoped but genuinely open-ended research tasks, not annotation. A given engagement might ask you to train a fine-grained image classifier under a hard parameter budget, distill a large teacher into a student that meets a latency target, produce a small diffusion model that hits a specific FID, adversarially train a classifier and honestly measure robust accuracy under AutoAttack, or post-train an open-weight LLM so it resists sycophancy and persuasion across multi-turn conversations without losing general capability. You are expected to own the loop end-to-end: build or synthesize the data, run the training, diagnose the loss curve that flattened, and write up what actually moved the number versus what didn't.

What the platform screens for

Mercor's screen is AI-led and follow-up heavy. Depth in at least one named area matters far more than breadth across all of them, and the follow-ups are designed to separate people who have run these experiments from people who have read about them. Expect to be pressed on specifics:

Concrete numbers from your own runs — dataset sizes, compute budgets, baseline vs. final metrics
Methodological hygiene: gradient masking, contamination, evaluation under a stated threat model, statistically sound comparisons
Trade-off reasoning (robustness vs. clean accuracy, compression vs. accuracy, per-language sampling temperature vs. transfer)
Framework fluency in PyTorch, JAX, or TensorFlow, and how you debug training when it fails silently
Credential floor: 3+ years of ML research experience with PhD work counting, plus a top-100 degree, FAANG-tier AI experience, or an equivalent publication or open-source record

Logistics

Fully remote and project-based. Hours are flexible and largely async, but experiments require real contiguous blocks — most researchers commit 10–20+ hours per week and some engagements scale toward full-time. Compute is typically provided by the client project. Pay is stated as observed on the platform, not guaranteed; rates vary by specialty scarcity and project scope.

What they look for

3+ years ML research experience (PhD research counts)
Deep expertise in at least one of: adversarial robustness, efficient CV, generative image modeling, LLM post-training, or multilingual pre-training
Production-level PyTorch, JAX, or TensorFlow
End-to-end training experience: data construction, training, diagnosis, evaluation
Rigorous evaluation practice — threat models, contamination control, sound comparisons
Top-100 degree, FAANG-tier AI experience, or equivalent publication/open-source record

Before you apply

Would you pass this screen?

Drawn from what this listing actually requires. Answered in your browser — nothing is sent to us, and we do not store your answers.

Do you have at least 3 years of machine learning research experience, including time spent on PhD research, published papers, or equivalent contributions?
Machine Learning Research Experience · required
Are you proficient in PyTorch, JAX, TensorFlow, or an equivalent deep learning framework at a level that allows you to write custom training loops and debugging code?
ML Framework Proficiency · required
Do you have a verifiable research track record through publications, significant open-source contributions, prior work at a FAANG or top AI company, or a degree from a top-100 university?
Domain Track Record · required
Have you trained a deep learning model from scratch and deployed it under real constraints (compute, model size, latency, or data limits)?
Hands-On Model Training · preferred

0 of 4 answered

Included with this listing

Interview prep — 11 questions

Matched to what this platform's screening actually measures. Answer each out loud in 60–90 seconds: claim → concrete example → implication.

What it tests: Validates hands-on end-to-end model training experience and troubleshooting ability; filters for researchers who've actually done the work versus read papers about it.
A strong answer: A concrete narrative covering data preparation, architecture choice with reasoning, hyperparameter decisions, at least one major training failure and how it was diagnosed (divergence, overfitting, loss spikes), and the final performance—the listener should understand both the technical moves and the judgment calls that shaped them.
Where people lose it: Describing a project at 10,000 feet ("we trained a classifier") without naming specific failure modes, debugging steps, or architectural decisions betrays lack of hands-on ownership; overly polished narratives suggest the work was primarily supervised rather than led.