Senior Research Scientist: Steerable Translation Models

DeepL; Greater London; Hybrid; GBP 120,000 - 180,000; Full time

Benefits offered by this job

Diverse and internationallyDistributed

Hybrid work

Virtual Shares

Regular in-person events

Hack Fridays

30 days annual leave

Competitive benefits

Job summary

DeepL is hiring a Senior Research Scientist to lead fine-tuning, post-training, and RL for its next-generation translation models. You will blend human expert data with synthetic data to steer models according to user instructions and context, working on models with hundreds of billions of parameters.

You will own the full lifecycle from prototyping to real-time production, establish robust evaluation, and mentor a fast-moving team of researchers and engineers in a hybrid UK-based setup.

Qualifications

• Experience making large models steerable and instruction-following.
• Hands-on expertise in post-training methods (SFT, DPO) and RLHF/RLAIF.
• Strong data-centric instincts for synthetic-data pipelines.
• Experience designing evaluation and reward signals with automated metrics and human-in-the-loop.
• Hands-on model training, experimentation, and production integration.

Responsibilities

• Lead fine-tuning, post-training, model-steerability, and RL for translation models.
• Fuse high-value human data with synthetic data to shape model behaviour.
• Build reward/models evaluators for translation and mitigate reward hacking.
• Drive models to ingest multimodal content to boost translation quality.
• Own lifecycle of model delivery from prototyping to production deployment.
• Establish evaluation, reproducibility, monitoring, and continuous model improvement.
• Mentor researchers and engineers to raise the bar for quality.

Skills

Steerable models

Instruction tuning

RLHF/RLAIF

Python

PyTorch

Communication

Mentoring

Tools

PyTorch

JAX

TensorFlow

DeepSpeed

Job description

DeepL is hiring a Senior Research Scientist to lead fine-tuning, post-training, and RL for its next-generation translation models. You will blend human expert data with synthetic data to steer models according to user instructions and context, working on models with hundreds of billions of parameters.

You will own the full lifecycle from prototyping to real-time production, establish robust evaluation, and mentor a fast-moving team of researchers and engineers in a hybrid UK-based setup.