ABOUT THE ROLE
We’re looking for experienced applied scientists in two tracks: pre-training, and post-training and continual learning. Each scientist will own a focused research direction within a track. The areas below describe the team’s research agenda; we are looking for depth in one or more of them.
The role involves designing experiments, writing the code to run them, analysing results and developing training methods that other teams can use. Scientists own the research questions, training objectives and evidence needed to decide whether a change works, and work with ML engineers to turn successful experiments into reliable training runs. We are a team of around 30 engineers with a flat structure, working directly with portfolio companies to define problems, evaluate models and bring successful work into production.
In the first three months, the expectation is to establish strong baselines in a chosen research area, develop and evaluate an improvement, and work with the relevant teams to test it in a downstream application.
TRACK 1: PRE-TRAINING
This track focuses on continued pre-training of open-weight models on commerce data, including how we represent that data, allocate training compute and transfer capabilities to smaller models. Research areas include:
Studying scaling behaviour, data mixtures and compute allocation to decide what goes into the next training run.
Developing mid-training methods, including data refinement, annealing schedules, context extension and alignment between text and semantic IDs for users, items and merchants.
Designing training sequences from user histories, merchant catalogues and market data, and studying how representations transfer across companies, languages and commerce categories.
Building synthetic data pipelines and measuring their contribution through controlled comparisons with real data.
Distilling the base model into smaller models, including through on-policy distillation, and measuring the trade-offs between quality, inference cost and latency.
Developing evaluations that help select pre-training runs and predict downstream performance, including checks for temporal leakage, generalisation to new users and items, and retention of existing capabilities.
Tabular foundation models are a possible specialisation: pre-training and in-context learning on structured commerce data, using synthetic and real tables and comparing results with established tabular methods.
Relevant experience includes continued pre-training or mid-training, data mixture research, scaling laws, tokenisation, entity representations, distillation, or tabular foundation models such as TabPFN and TabICL.
TRACK 2: POST-TRAINING AND CONTINUAL LEARNING
This track focuses on adapting LCM to commerce tasks and learning from interactions after deployment. The goal is to turn signals from a continuous stream of billions of user sessions across the portfolio into regular updates to model weights, while retaining the capabilities the model has already learned. Research areas include:
Developing post-training methods using supervised fine-tuning, preference optimisation such as DPO, and reinforcement learning with verifiable or learned rewards, including GRPO and PPO.
Learning from session histories, searches, clicks, purchases, repeat orders and operator feedback. This includes data selection, credit assignment and reward design that account for exposure bias, delayed outcomes and the difference between observed engagement and incremental business value.
Studying how to combine new interactions with historical data, how frequently to update models, when an update is ready for deployment, and how to preserve performance as user behaviour and catalogues change.
Building RL environments with simulated users, catalogues and multi-step commerce tasks, including inspecting trajectories, validating rewards and testing whether improvements in simulation transfer to real interactions.
Developing reward models and LLM-based graders, calibrating them against expert judgments and observed outcomes, and testing for reward hacking, sycophancy and excessive concentration on popular items.
Training decision models for rankings, offers and promotions using contextual bandits or offline RL, with off-policy evaluation to estimate value and uncertainty, identify gaps in the available evidence and inform controlled online experiments.
Developing automated experiments that propose and test changes to training methods, prompts, tools or agent control flow, evaluated on held-out tasks and regression tests before acceptance.
Relevant experience includes language-model post-training, reinforcement learning, reward modelling, agent training environments, continual learning, or offline RL and bandits applied to real decisions. The training experience required will depend on the research direction.
RESPONSIBILITIES ACROSS BOTH TRACKS
Define success with portfolio teams and evaluate models against strong baselines, including existing production systems, accounting for inference cost, latency and the business outcome the model is meant to improve.
Design reproducible experiments with appropriate ablations, uncertainty estimates and failure analysis, using future-period and held-out evaluations, and test whether offline improvements hold online.
Work with data and engineering teams on provenance, permitted use across companies and privacy-aware handling of customer histories.
Publish research and contribute models, benchmarks, environments and training methods to open source where the work and data can be shared.
Mentor colleagues, review experiment designs and communicate findings clearly enough to support training and deployment decisions.
MINIMUM QUALIFICATIONS
A PhD or equivalent industry research experience in machine learning, computer science, statistics or a related field.
At least five years of research or applied science experience in machine learning, including doctoral research, with hands-on experience designing and running model-training experiments. For Track 1, this should include pre-training or mid-training of language models. For Track 2, it should include language-model post-training, or reinforcement learning and decision-model training relevant to the specialisation.
Evidence of original research through papers, released models, public benchmarks, open-source projects or significant research contributions to deployed systems.
Strong Python and PyTorch skills, with experience implementing experiments and debugging training runs in a shared research codebase. Work on large language models requires hands-on experience with distributed training.
Sound experimental judgment, including strong baselines, statistical analysis and the willingness to report negative results.
Clear written and verbal communication.
PREFERRED QUALIFICATIONS
Experience implementing and debugging training methods in frameworks such as Megatron-LM, DeepSpeed, NeMo RL, verl or OpenRLHF, depending on the chosen track.
Research in recommender systems, generative retrieval, semantic IDs or user behaviour modelling.
Experience with synthetic data, model distillation or compression for production.
Experience working with marketplace or commerce data and evaluating changes through online experiments.
WHAT WE OFFER
Autonomy to design experiments and make technical decisions within your research direction.
Proprietary commerce data, B300 and H200 GPUs for large-scale training, and external models for evaluation and benchmarking.
The opportunity to take research from early experiments through publication and deployment, working directly with portfolio teams on products used by millions of people.
Colleagues who have released Hugging Face models with more than 100 million downloads, authored papers at NeurIPS, created benchmarks and built multiple production AI systems.
An uncapped budget to recruit and mentor research interns for your projects, with access to candidates through our connections with leading universities.
Current ML frameworks and unlimited access to coding assistants.
A base at our Amsterdam office and AI House, with relocation support. We also consider remote arrangements for candidates with particularly strong experience relevant to the role.
Competitive compensation and a top-spec MacBook Pro.
HOW TO APPLY
If you want to work on these problems, we would like to hear from you. Please indicate the track and research areas that best match your experience, and include examples of work that show your contribution.
Why Prosus?
At Prosus, your colleagues include 60,000 AI agents… and your job is to figure out what that actually means. We're one of the largest operators of agentic AI on the planet, inside a $100B+ portfolio across 100+ companies and two billion people. The humans here own what AI can't do alone: judgment, creativity, and the call on what gets built next.
Own the impact. Work the ecosystem. Earn your growth. Build with backing.
Curious about our perks and benefits? Take a look at this page https://prosusbenefits.com/.
Our Diversity & Inclusion Commitment
We respect the dignity and human rights of individuals and communities wherever we operate in the world. Building an inclusive workplace where everyone feels welcome and can thrive is critical for us. We provide access to education, which helps everyone understand the important role they play and the positive impact they can have.
Want to know more about life at Prosus? Explore our culture and values https://www.prosus.com/the-prosus-way.