Machine (Meta) Learner

United States Digital Space LLC; Heidelberg; Vor Ort; EUR 90.000 - 130.000; Vollzeit

Benefits dieser Stelle

VSOP equity

30 days of paid holiday per year

Statutory social insurance

Conference travel and learning

Flexible hybrid work

High-end laptop and compute access

Zusammenfassung

United States Digital Space LLC seeks a research scientist to advance causal, reasoning-first models with a focus on PFNs, meta-learning, and priors that determine learnable capabilities. You will form hypotheses, build test systems, and produce reproducible research, open-source work, and production-relevant capabilities.

Responsibilities include designing priors and synthetic task distributions, developing temporal and goal-conditioned models, and building rigorous evaluations with baselines

Qualifikationen

• PhD in ML, physics, statistics or related field (or equivalent research experience).

Aufgaben

• Shape and pursue research questions around PFNs, meta-learning, and in-context learning.

Kenntnisse

PFNs

Meta-learning

Bayesian inference

Neural Processes

Representation learning

Causality

Active learning

Python

PyTorch

JAX

Ausbildung

PhD in ML/Physics/Statistics or related field

MSc with exceptional experience

Tools

Python

PyTorch

JAX

PyTorch Lightning

Weights & Biases

Docker

AWS

Jobbeschreibung

At the company, we build causal, reasoning-first models that learn from a handful of examples and adapt without retraining. We are looking for a research scientist to advance the foundations of that approach, with a particular focus on Prior-Data Fitted Networks, meta-learning and the priors that determine what our models can learn. This is a research role with real implementation responsibility. You will form hypotheses, build the systems needed to test them and turn strong results into reproducible research, open-source work and production-relevant capabilities.

Tasks

Our research revolves around synthetic world data, deep-learning models trained and validated against it, and capable embedders across domains and modalities. You will:

• Shape and pursue research questions around PFNs, meta-learning, in-context learning, representation learning, causality, active learning and adaptive decision-making.
• Design priors and synthetic task distributions that expose models to useful structure, uncertainty and failure modes.
• Develop model architectures and training methods for temporal, goal-conditioned and dynamical settings.
• Build rigorous evaluations, including strong baselines, ablations, calibration tests and out-of-distribution diagnostics.
• Implement research ideas reliably in Python and PyTorch, and improve the data and experiment pipelines around them.
• Contribute to top-tier publications, open-source releases and the wider research agenda at the company.

Requirements

We are looking for research scientists with a strong background in one or more of:

• Deep expertise in PFNs, meta-learning, Bayesian inference, Neural Processes, representation learning, causality, active learning or a closely related area.
• A record of generating original research hypotheses and testing them with scientific rigor.
• Strong experimental judgment: you can distinguish optimization failure, prior misspecification and distribution shift.
• Reliable implementation skills in Python and PyTorch or JAX.
• A PhD in machine learning, physics, statistics or a related field, or equivalent research experience.
• The ability to work independently, explain difficult ideas clearly and change your mind when the evidence demands it.
• We are primarily hiring at senior level. We are also open to exceptional candidates with fewer years of experience who can demonstrate comparable depth, judgment and ownership.

Recommended qualifications

• A PhD in ML, Physics, or equivalent — or an MSc with exceptional experience
• A strong grasp of causality, meta-learning, PFNs, and active inference
• The ability to work independently and think from first principles
• Hands‑on experience with modern ML tooling (Python, PyTorch) and research workflows
• An outcome‑oriented mindset

Nice to have

• Causal modeling, active learning or Bayesian optimization.
• Reinforcement learning, control, time‑series modeling or dynamical systems.
• Synthetic-data generation, graph‑based models or simulation environments.
• Publications at NeurIPS, ICML, ICLR or comparable venues.
• Meaningful open‑source contributions.

Benefits Where This Can Go

You will help define the company's research agenda, not just execute it. As the team grows, there is room to lead a research direction, mentor incoming scientists, and shape how our published work and open‑source contributions reach the wider community. And as the company begins working with its first customers, the research you do here is increasingly likely to leave the lab and reach real-world deployment.

Our Culture

• are scientists at heart, with a builder's mindset,
• are open to challenge, grounded in curiosity and respect,
• welcome diverse perspectives and value thoughtful, open debate,
• focus on outcomes and real-world impact,
• foster an environment of support, inspiration, and freedom for everyone to do their best work.

Perks & Benefits

• VSOP equity: a real stake in what we build.
• 30 days of paid holiday per year.
• Statutory social insurance.
• Conference travel and role-relevant learning.
• Flexible hybrid work, with roughly one in-person team meet-up per month.
• A high-end laptop and access to the compute required to do serious research.

Tools and Infrastructure

• Python, PyTorch, and PyTorch Lightning
• Weights & Biases and reproducible experiment workflows.
• Docker, AWS, RunPod and comparable cloud infrastructure.

Send us your CV or LinkedIn along with any selected works or contributions.