Member of Technical Staff, Model Training (post-training research)

What you'll do

This is the research-heavy member of the Model Training trio. You study how learning algorithms, training data and evaluation methods interact to produce better reasoning, generalization and agent behaviour, then turn what works into scalable training methods. The ad frames it as highly experimental: you form hypotheses, design ambitious experiments and build the methods yourself.

The listed work covers SFT, reinforcement learning and on-policy distillation across large language and multimodal models; learning objectives, optimization dynamics, exploration and credit assignment; training recipes that jointly tune data mixtures, sampling and curricula; and learning from demonstrations, human preferences and corrections. A distinctive thread is the co-evolution of generators, reward models and graders, which is where human evaluation meets model training. You own initiatives end to end, including ablations and large-scale validation. The role is full-time and remote.

What you need

• Strong foundations in machine learning, deep learning, optimization and modern training techniques
• Hands-on experience training, fine-tuning or post-training large language or multimodal models
• Deep expertise in at least one of RL, distillation, reward modeling, data selection or learning dynamics
• Strong Python and PyTorch, enough to implement, debug and optimize complex training pipelines
• Experience designing controlled experiments and turning results into research directions
• Independence on ambiguous problems, with quick iteration on experimental evidence
• A rigorous, reproducibility-minded approach

Nice to have

• RLHF, RLAIF, preference optimization or RL for reasoning and agentic tasks
• Distributed training, large-scale experimentation and GPU optimization
• Reward models, verifiers, graders or automated evaluation systems
• Publications, open-source projects or production systems in post-training or RL
• Synthetic data generation, curriculum learning or iterative self-improvement

What it pays

The compensation notice inside the ad gives a national base salary range of $200,000–$320,000 for the full-time position, and that is the figure on this page, with equity for all employees and possible performance bonuses depending on role. micro1's card shows a wider $600,000–$1,800,000/year. The detail panel also shows an hourly figure of $7–8 that matches neither, so treat it as a display error and ask about the real package early. The notice lists up to 100% reimbursement of health-insurance premiums, paid time off and a 401(k) with company match, which makes this an employee-style package rather than the contractor engagement most micro1 listings carry.

Worth knowing

Good:

• Research scope with real room to set direction, not only run someone else's experiments
• The reward-model and grader angle suits people who already think hard about evaluation quality
• Full-time, remote, with a stated benefits package

Less good:

• The pay figures on the listing disagree with each other, so confirm the real package early
• A single opening and a research-grade bar; publications or equivalent evidence will likely matter
• The ad does not say what compute you would have or how research is prioritized
• Hiring uses AI tools to assist screening, as the ad itself discloses

About this listing

Read on micro1's jobs board on 28 October 2026. The three Model Training listings (infra and platform, applied research, post-training research) share one pay line and one compensation notice but have different scopes; this page covers the post-training research role. The compensation notice, benefits and every requirement above are the ad's published text. See micro1.