Thomson Reuters

Lead Research Engineer, Large-Scale Agentic RL Training

Thomson Reuters
Apply
2 days ago
Zug, Switzerland or London, United KingdomStaff+

Responsibilities

  • Own the architecture and technical evolution of the large-scale asynchronous RL training stack.
  • Profile and optimize the full RL loop, including generation, tool calls, reward and judge computation, training, and weight synchronization.
  • Design fault-tolerant pipelines for models interacting with tools and agent harnesses during training.
  • Diagnose numerical, distributed, and performance issues across training frameworks, inference engines, and model code.
  • Partner with Research Scientists to test algorithms, rewards, and environments quickly at scale.
  • Set standards for code quality, testing, reproducibility, and technical execution.
  • Mentor engineers through code reviews, technical guidance, and skills development.

Requirements

  • MSc or PhD in computer science, machine learning, or a related field, or equivalent industry experience.
  • At least 3 years of professional experience in ML engineering or research engineering involving large-scale model training.
  • Advanced PyTorch experience and extensive experience training large models on multi-node GPU clusters with Megatron-LM/Megatron-Core, FSDP/DTensor, or DeepSpeed.
  • Deep understanding of tensor, pipeline, context, and expert parallelism and their memory and communication trade-offs.
  • Hands-on experience with vLLM or SGLang, including serving large Mixture-of-Experts models.
  • Strong GPU profiling and optimization skills involving memory, NCCL communication, CUDA graphs, and kernel bottlenecks.
  • Strong software engineering skills and ability to work in large open-source codebases.
  • Proven ability to lead complex technical projects end to end and mentor engineers.
  • Experience diagnosing numerical, distributed, and performance issues and maintaining code quality, testing, and reproducibility.
  • Preferred experience with NeMo-RL, verl, OpenRLHF, or TRL, especially asynchronous or non-colocated designs.
  • Preferred experience training or serving models above 100 billion parameters at 100k+ token contexts.
  • Preferred experience with Ray, Kubernetes, or Slurm on AWS, Azure, or Google Cloud.
  • Preferred experience with numerical-parity testing and contributions to vLLM, Megatron, PyTorch, or similar open-source projects.
  • Publications or technical reports at ML or systems venues such as MLSys, NeurIPS, or OSDI are preferred.

Benefits

  • Flexible hybrid work model for office-based roles.
  • Work from anywhere for up to 8 weeks per year and additional flexible work-life policies.
  • Career development, continuous learning, and Grow My Way programming.
  • Comprehensive benefits including flexible vacation, two company-wide Mental Health Days, Headspace access, retirement savings, tuition reimbursement, incentive programs, and wellbeing resources.
  • Two paid volunteer days annually and opportunities for pro-bono consulting and ESG initiatives.
  • Inclusive culture with reasonable workplace accommodations and equal employment opportunity.

Tech Stack

Categories

Thomson Reuters

About Thomson Reuters

10,000+ employees

Thomson Reuters builds research platforms, workflow software, and data services for legal, tax and accounting, compliance, and government professionals, and operates the Reuters global news service. Flagship products include Westlaw for legal research and ONESOURCE and Checkpoint for tax and accounting, sold primarily via subscriptions and enterprise licenses. A public company headquartered in Toronto, it was formed in 2008 by combining Thomson Corporation and Reuters Group and is listed on the TSX and Nasdaq as TRI.

Contact me