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

What you'll do

This is closer to contract research than to annotation. You get well-scoped but open-ended empirical ML problems and solve them by actually training models: image classifiers and generative image models from scratch, and fine-tuning of open-weight language models. The recurring theme is doing a lot with a little (limited data, compute and parameter budgets) and making models robust, both to adversarial inputs and to adversarial conversations.

The ad names five specialisms; you need strength in one or more:

• Adversarial robustness: PGD-based or TRADES adversarial training, evaluation under standard threat models such as AutoAttack, avoiding gradient masking, handling the robustness and accuracy trade-off
• Efficient computer vision: fine-grained classification with few examples per class, quantisation, pruning and distillation, deployment under size or latency limits
• Generative image modelling: diffusion, GANs, VAEs or flows trained from scratch, iterating against FID, getting good samples from small models quickly
• LLM post-training and behavioural robustness: SFT and preference optimisation (DPO, RLHF, RLAIF), building your own datasets, resisting sycophancy and persuasion over multiple turns, calibrated confidence, targeted behaviour changes that keep general capability
• Multilingual pre-training: low-resource languages from scratch, tokeniser design across scripts, balancing very unequal per-language data

Scaling laws, curriculum learning, benchmark construction and contamination control, calibration and synthetic data are listed as pluses.

Who fits

• 3+ years of ML research experience, with PhD research counting toward it
• Strong PyTorch, JAX, TensorFlow or similar
• A degree from a top-100 university, FAANG or comparable AI-company experience, or an equivalent record through publications or strong open-source work

The alternative routes in that last point matter: a solid open-source training record can stand in for pedigree.

What it pays

$100–120 per hour, paid weekly on Stripe or Wise. The band is narrow, so the ad's figure is a good guide to what you will actually be offered.

Worth knowing

Good:

• Real research work, training models end to end
• PhD years count toward the experience bar
• Open-source contributions accepted in place of a top-100 degree or FAANG history
• Remote, flexible, project-based

Less good:

• No hours, duration or client disclosed
• Compute access is not mentioned; ask what hardware you will train on before you accept
• Projects can be extended, shortened or ended early
• H-1B and STEM OPT holders are excluded
• Contractor terms, no benefits

About this listing

Posted by Mercor as a remote hourly contract, read on 24 September 2026. The pay band, specialisms, qualifications and payment terms are the ad's own. See Mercor.