LLM Research Scientist – Pre-training, Computer Vision, Adversarial Robustness

This is a fully remote position, open to applicants in United States.

📋 Description

• Develop and train image classifiers and generative image models from the ground up.

• Fine-tune open-weight language models to enhance performance.

• Optimize models while adhering to constraints related to data, compute resources, and model size.

• Enhance model robustness against adversarial inputs and challenging conversational scenarios.

• Compress models to satisfy stringent size and latency requirements without sacrificing accuracy.

• Diagnose and troubleshoot training-related issues effectively.

• Engage in empirical and exploratory machine learning research.

• Collaborate with top AI researchers on impactful and challenging projects.

• Participate in initiatives focused on training and improving cutting-edge AI systems.

⛳️ Requirements

• Minimum of 3 years of experience in machine learning research; PhD research is applicable.

• Strong proficiency in PyTorch, JAX, TensorFlow, or other similar machine learning frameworks.

• Degree from a top-100 university, experience at a FAANG or similar AI organization, or a comparable research background demonstrated through publications or significant open-source contributions.

• Expertise in one or more areas such as adversarial robustness, efficient computer vision, generative image modeling, post-training for LLMs, behavioral robustness, or multilingual pre-training.

• Experience in adversarial training of image classifiers, evaluating robust accuracy, developing threat models, and understanding the trade-offs between robustness and accuracy.

• Proven experience in end-to-end training of image classifiers, model compression, and deployment under stringent size or latency constraints.

• Familiarity with training diffusion models, GANs, VAEs, or flow-based generative models from inception.

• Experience with supervised fine-tuning, DPO, RLHF, RLAIF, synthetic data generation, weak supervision, and rejection sampling.

• Expertise in shaping conversational behaviors, including resistance to persuasion, calibrated confidence, and acceptance of corrections.

• Background in training multilingual or low-resource language models from scratch, tokenizer design, and cross-lingual data balancing.

• Experience in scaling laws, curriculum learning, model evaluation, uncertainty estimation, calibration, data augmentation, or synthetic data for robustness is advantageous.

• Must be capable of working as an independent contractor.

• Mercor is currently unable to support H1-B or STEM OPT candidates.

🏝️ Benefits

• Flexible, project-based working arrangements.

• Competitive salary package.

• Fully remote working environment.

• Freedom to set your own working hours.

• Weekly payments via Stripe or Wise.

• Project durations may be adjusted based on needs and performance.

• Reasonable accommodations available upon request.

• Referral bonuses of up to $480 for each successful referral, subject to certain limits.