← Back to All Openings

Senior Research Scientist – Hierarchical Belief‑Aware Variational Bottleneck

corpora-jobs-1778796293285-db9d41c6 - Frontier Development
Research ScientistSenior1 position
⚡

Why This Role is Different

Frontier Development Role

You will pioneer a new class of belief‑aware abstractions that fuse information‑theoretic regularization with hierarchical policy decomposition. Your work will directly tackle the hardest credit‑assignment bottleneck in decentralized RL, enabling agents to reason about uncertainty at multiple temporal scales.

The Frontier Element

By embedding a variational bottleneck in belief space, you will create the first end‑to‑end differentiable pipeline that learns to discard spurious observations while preserving essential coordination cues—an approach that has never been demonstrated at scale in MARL.

🔬

Project Context

Research Area

Hierarchical Belief‑Aware Abstraction

From: Partial Observability Amplification of Misalignment

Why This Role is Critical

To design and train a multi‑scale belief hierarchy that compresses sensory embeddings through a variational bottleneck conditioned on observation history and a shared world‑model prior, thereby preserving task‑relevant modalities and reducing credit‑assignment errors in partially observable MARL.

What You Will Build

A modular belief‑hierarchy framework, training pipelines, evaluation metrics, and integration modules that expose belief divergence signals to the rest of the BAAC stack.

🛠

Key Responsibilities

  • Design and implement a multi‑scale belief hierarchy with a KL‑regularized variational bottleneck.
  • Integrate the hierarchy with a shared world‑model prior and train end‑to‑end on large‑scale MARL benchmarks.
  • Develop diagnostics to quantify task‑relevant versus spurious latent factors and publish findings.
  • Prototype real‑time inference on edge hardware to validate bandwidth and latency constraints.
  • Collaborate with the communication and JBWM teams to expose belief divergence signals for downstream modules.
🎯

Required Skills & Experience

Technical Must-Haves

Variational Autoencoders and Information‑Theoretic Regularization

Expert
Designing KL‑constrained belief bottlenecks that preserve task‑relevant information.

Multi‑Agent Reinforcement Learning (CTDE & Dec‑POMDP)

Advanced
Understanding credit‑assignment dynamics under partial observability.

Graph Neural Networks and Attention Mechanisms

Advanced
Encoding relational structure among agents within the belief hierarchy.

Python, PyTorch, and CUDA

Proficient
Rapid prototyping and GPU‑accelerated training.

Experience Requirements

  • 5+ years of research in RL or representation learning with at least 3 peer‑reviewed conference papers.
  • Track record of publishing in top venues (NeurIPS, ICML, ICLR, AAMAS).

Education

PhD in Computer Science, Electrical Engineering, or Machine Learning with a focus on RL or probabilistic modeling.

⭐

Preferred Skills

  • Experience with causal discovery or hierarchical option learning.
  • Simulation expertise in SMAC, MPE, or UAV swarm environments.
  • Knowledge of object‑oriented world‑modeling frameworks.
🤝

You Will Thrive Here If...

  • Thrives in high‑autonomy environments and is comfortable shipping prototypes before formal reviews.
  • Demonstrates a passion for pushing theoretical ideas into production‑ready systems.
📈

Impact & Growth

12-Month Impact

Within 12 months, deliver a benchmark‑competitive BAAC agent that reduces credit‑assignment error by ≥30 % on SMAC and MPE, open‑source the belief‑hierarchy code, and publish a paper on belief‑aware abstraction.

Growth Opportunity

Lead a growing research team of 5–7 scientists, shape the product roadmap for BAAC, and transition from prototype to production‑grade deployment.

Ready to Push the Boundaries?

If this sounds like the challenge you have been looking for, we want to hear from you. We value what you can build over where you have been.