Research Scientist / Engineer, Alignment & Reward Modeling (all genders)

ellamind • Berlin, KY

4d • Remote

About The Position

ellamind is hiring a Research Scientist / Engineer to develop alignment and reward-modeling methods for its language models. The role focuses on building reward models and LLM judges for tasks that require rubric-based, generative, and process-level assessment, and validating those judges against careful human judgment. The successful candidate will train models to be honest about their actions and limitations and build evaluations that identify unreliable behavior.

The role is embedded in RL training, with its work feeding directly into model rewards. It includes translating requirements for regulated work into training signals, investigating reward hacking and specification gaming, and running alignment experiments during RL post-training. The work also involves assessing the effects of these methods on model capability and efficiency, and developing safety and reliability evaluations informed by real enterprise and public-sector cases.

The position can be performed from anywhere in the EU, ideally Germany. ellamind offers access to compute across EuroHPC and commercial providers, an office option in Bremen or Berlin, and opportunities to contribute to public work, including a planned open-source release in Q1 2027.

Requirements

• Experience training or evaluating LLMs and hands-on experience in at least one of reward modeling, RLHF or preference learning, LLM-as-judge, or alignment or safety evaluation.
• Strong Python and PyTorch skills sufficient to independently take an experiment from idea to result.
• A critical approach to metrics and an ability to consider how apparently good results could be misleading.
• Ability to translate vague behavioral problems into measurable and trainable criteria.
• Clear writing and communication skills.
• Residence in the EU and authorization to work there.
• English proficiency; English is sufficient for the role.

Nice To Haves

• Experience with interpretability or scalable oversight.
• Publications on reward modeling or evaluation.
• Experience in a regulated field.
• Familiarity with the EU AI Act.
• German language skills.

Responsibilities

• Build reward models and LLM judges for rubric-based, generative, and process-level tasks that cannot be assessed by unit tests alone.
• Validate judges against careful human judgment.
• Turn behaviors important in regulated work—such as following policy and instructions, asking for clarification, stopping instead of guessing, accurately reporting tool results, and expressing appropriate confidence—into training signals.
• Investigate reward hacking and specification gaming, study how judges fail, and develop fixes.
• Run alignment experiments within RL post-training using preference data, synthetic critiques, and principle-based feedback.
• Assess and report the capability and efficiency costs of alignment experiments.
• Build safety and reliability evaluations grounded in real-world enterprise and public-sector cases.
• Communicate research findings clearly.

Benefits

• Work from anywhere in the EU, ideally Germany.
• Access to hundreds of GPUs across EuroHPC and commercial compute.
• Use of AI tools that support the work without a procurement process.
• Opportunity to have your name on public work, including a planned open-source release in Q1 2027.
• Direct access to the founders and a research team to learn from.
• Optional offices in Bremen and Berlin; the Bremen office is by the water and offers free drinks and smoothies.
• Competitive pay and equity options.

What This Job Offers

• Job Type: Full-time
• Career Level: Senior
• Education Level: No Education Listed