Research topics:

  • AI safety & alignment
  • Large language models
  • Pretraining at scale
  • RL posttraining at scale
  • Synthetic-data generation
  • Model evaluation & red-teaming

We are hiring AI research engineers to support our research efforts on aligning large language models from token zero and to put this research into practice with the next generation of Apertus, one of the world’s largest fully-open LLM programs and the largest pretraining effort in the non-commercial space.

The positions are funded by a generous $1M award from Coefficient Giving and hosted jointly by

Most alignment today is postprocessing tacked onto a model whose representations have already crystallized during pretraining—what we call “lipstick-on-a-pig alignment”. We take a different view. We raise models: value formation is woven throughout the entire training process, starting from token zero in pretraining, so that alignment is tied to capability rather than layered on top of it and so that alignment survives intensive RL posttraining (high-level vision here). Our early results are encouraging.

As an AI engineer on this project, you’ll apply these methods by training and evaluating LLMs at various scales, and by translating them into Apertus’s production LLM pipeline scaling to thousands of GPUs, designing and applying production recipes for a new 700B model trained from the ground up starting in 2026.

Areas of expertise

These areas overlap, and we don’t expect any single person to cover them all. Apply based on your strengths and tell us which fit you best.

  • Pretraining. Our goal is to instill values from pretraining token zero, which only works if the large runs underneath stay fast, stable, and healthy. Relevant expertise: distributed training, Megatron-LM integration, parallelism and throughput optimization, training-stability debugging at scale, scaling model size (MoE and dense).
  • RL & posttraining. Safety pretraining is worthless if it breaks under intensive RL posttraining, so we need a robust posttraining stack. Relevant expertise: RLHF/RLVR/RLAIF and preference-optimization pipelines, reward modeling, rollout-generation infrastructure, on-policy distillation, agentic and multi-agent training.
  • Data infrastructure. Our methods rely heavily on synthetic data to convey values to models, so the pipeline that produces and curates it is crucial. Relevant expertise: synthetic-data generation, teacher-LLM orchestration, web-scale data curation and filtering (dedup, classification, quality/toxicity), efficient inference, quality control.
  • Evaluation & red-teaming. A safety method is only as trustworthy as our ability to measure whether it holds, reproducibly and across checkpoints. Relevant expertise: LLM-as-a-judge infrastructure, inference optimization, reproducibility and ablation hygiene, automated jailbreak and persona-drift testing, interpretability tooling.

What we offer

  • Serious compute. Secured access to the Swiss National Supercomputing Centre’s Alps cluster (10,000+ GH200 GPUs), with a budget exceeding 10M GPU-hours per year for pretraining.
  • Real production impact. Methods that pass validation feed directly into the next Apertus training run. You’ll build a real production LLM, not merely a prototype that might one day matter.
  • Highly competitive compensation, well above standard academic scales.
  • Paper co-authorship. While your focus is on engineering, you’ll also contribute as a co-author to the research papers that the team will publish.
  • An open, international research environment, an extremely well-funded national research system, generous travel support, and an office next to a stunning lake and even more stunning mountains.

Qualifications

  • Master’s or PhD degree in computer science, engineering, or a related field.
  • Relevant engineering experience with deep learning in general, and LLMs in particular. Having shipped systems, not only prototypes, is a plus.
  • Strong programming skills and hands-on experience with modern ML frameworks. Depending on area: distributed/large-scale training, RL and posttraining pipelines, data infrastructure, or evaluation and interpretability tooling.
  • A pragmatic, quality-obsessed engineering sensibility, and genuine interest in the mission of safe, open AI.

Time frame

Positions are for two years, with an ideal start in 2026. There is some flexibility on start date.

About EPFL

EPFL ranks among the world’s top universities in computer science. It is located in Lausanne, Switzerland, a vibrant, highly international city in an Alpine setting on the shores of Lake Geneva, in the heart of Europe. English is the main language spoken at EPFL, and no French is required.

How to apply

To apply, please complete this short online form.

Review of applications begins immediately and continues until the positions are filled.

Contact: