Applied Scientist

Applied Researcher, Post-Training

Location: In-person - 3 days in-office in NYC or Seattle

Employment Type: Full-time

Focus: Reinforcement Learning, Post-Training, Training Data, Evals, Frontier AI

Profile: Research-first applied researcher with deep hands-on RL experience

About Our Client

Our client works directly with frontier AI labs on some of their hardest post-training problems.

As demand from advanced research customers continues to grow, the team is hiring applied researchers who can help design better environments, datasets, reward signals, evaluation harnesses, and post-training loops for frontier models.

This is a research role first. The team cares most about depth of understanding in reinforcement learning, post-training, and training data quality. Strong engineering ability is useful, but the core bar is research judgment, taste, and the ability to improve model behavior through better data and training signals.

About the Role

Our client is hiring an Applied Researcher, Post-Training to work directly with customer research labs and help turn ambiguous research goals into concrete experiments and deliverables.

This person will design and run post-training loops, inspect high-fidelity trajectories, build reward functions and eval harnesses, and determine which signals actually improve model capability.

The ideal candidate has done real, hands-on reinforcement learning work — at a frontier lab, AI research lab, data lab, or through a serious independent project — and has results to show for it.

What You’ll Do

• Work directly with customer research labs to understand their hardest post-training problems
• Turn unclear research goals into concrete environments, datasets, experiments, and deliverables
• Design and run post-training loops, including reinforcement learning, preference optimization, and supervised fine-tuning
• Work with methods such as GRPO, RLHF-style training, preference data, reward modeling, and related post-training approaches
• Figure out which signals actually improve model performance and which ones only look good on paper
• Shape, inspect, and improve high-fidelity trajectories and training data
• Review data closely for correctness, edge cases, realism, and overall quality
• Build reward functions, rubrics, validators, and evaluation harnesses
• Design evals that hold up against reward hacking and show real capability gains
• Read post-training signals and experiment results across distributed GPU clusters
• Decide what experiment, dataset, reward signal, or training adjustment to try next

What We’re Looking For

• Deep hands-on experience with reinforcement learning or post-training
• Strong understanding of training data quality and how data impacts model behavior
• Experience running or contributing to real post-training experiments
• Experience with reinforcement learning, preference optimization, SFT, reward functions, evals, or model behavior research
• Ability to reason from messy experiment results and identify what to try next
• Strong taste around trajectories, tasks, environments, and training data quality
• Ability to identify edge cases, reward hacking, bad incentives, and misleading evaluation results
• Comfort working directly with research customers on ambiguous technical problems
• Evidence of real results from frontier lab work, AI lab work, data lab work, or serious independent research projects
• Strong communication skills and ability to translate research goals into concrete execution

Bonus Experience

• Experience with GRPO or related reinforcement learning methods
• Experience building reward functions, validators, eval harnesses, or rubrics for model training
• Experience working with distributed GPU training environments
• Experience with RLHF, RLAIF, preference optimization, synthetic data generation, or long-horizon task environments
• Experience at a frontier AI lab, emerging AI lab, post-training data company, or applied AI research team
• Independent research projects with public artifacts, measurable results, or strong technical depth
• Experience working directly with external research teams or frontier lab customers

Who Will Thrive Here

• A researcher who has actually trained or post-trained models, not just read the papers
• Someone who deeply understands reinforcement learning signals and training data quality
• A person with strong data taste who can tell when a trajectory, reward function, or eval is misleading
• A researcher who enjoys working close to frontier lab customers
• Someone who can operate in ambiguity and turn vague research goals into concrete experiments
• A high-slope builder who cares about real model improvement, not vanity metrics

Why This Opportunity

• Work directly with frontier AI labs on real post-training problems
• Join a small applied research team with growing customer demand
• Focus on the highest-leverage parts of post-training: data, signals, evals, rewards, and experiments
• Help define what good post-training data and environments look like for frontier models
• Work on research problems where judgment, taste, and hands-on RL experience matter
• Build systems and processes that help models gain real capabilities, not just better benchmark numbers

Ideal Candidate Profile

The ideal candidate is an applied researcher with a genuine spike in reinforcement learning and post-training.

They have hands-on experience designing experiments, working with training data, building reward signals, inspecting model behavior, and figuring out what actually improves performance. They are research-first, deeply technical, and comfortable working directly with frontier AI customers on ambiguous problems.

This person should care about model capability gains, data quality, and research taste more than traditional software engineering scope.