About the role:

Cobalt is seeking researchers and engineers with direct experience in language model post-training and evaluation, to produce the expert reasoning and evaluation data frontier labs use to improve model behavior.

This opportunity is suited to people who have worked on the parts of the stack closest to how a model actually behaves: supervised fine-tuning, preference optimization and RLHF, reward modeling, inference-time reasoning methods, and the design of evaluations that hold up. You may have done this in a lab, in industry, or in serious open-source work.

You do not need prior experience in data annotation. What matters is that you understand why models fail in the ways they do, and that you can write the kind of data and criteria that fix it.

What you'll do:

Depending on the project, you may:

Produce written reasoning traces and expert reference answers on hard technical problems, at the standard of quality a post-training set requires rather than merely correct answers
Author evaluation items and benchmark tasks with verifiable success criteria, including cases specifically designed to separate genuine capability from pattern matching
Evaluate and rank model outputs on technical questions, articulating precisely what separates a strong response from one that is fluent but subtly wrong
Design rubrics, reward criteria, and partial-credit schemes for multistep tasks, and identify where a criterion would be gameable or would reward the wrong behavior
Classify observed failures into a consistent taxonomy, and assess whether a stated conclusion is supported by the underlying reasoning

Projects follow their own guidelines, formatting conventions, and quality standards, and you will work with feedback from reviewers and lab research teams.

Required qualifcations

Direct experience with language model post-training or evaluation, such as supervised fine-tuning, preference optimization or RLHF, reward modeling, or benchmark and eval design, whether in research or industry
A PhD in a quantitative discipline, or equivalent depth demonstrated through published work, open-source contributions, or production systems
Strong coding ability in Python, and working command of at least one deep learning framework
Understanding of common failure modes in current models, including reward hacking, sycophancy, and answers that are right for the wrong reasons
Ability to explain each step of your reasoning clearly in writing, and to specify criteria precisely enough that another annotator would apply them the same way

Why join Cobalt AI:

Advance frontier AI where it counts. Apply your expertise to data that frontier labs cannot obtain any other way, where your reasoning directly shapes how the next generation of models works through technical problems.
Grow professionally. Expand your influence through evaluation projects, advisory roles, and research collaborations, while developing a working understanding of how frontier models are trained and assessed.
Work with a top-tier network. Collaborate with researchers and engineers from leading institutions and labs on high-impact, flexible work.
Set your own schedule. Flexible 10 to 40 hour weeks that fit around your existing work and your life.
Competitive pay. Rates vary by project and are determined by a number of factors, including scope, skillset, and experience.

Cobalt

New Jobs

LLM Post-Training and Evaluation Researcher (Contract)