# Research — Paulius Rauba

Canonical URL: https://www.pauliusrauba.com/research

I am a Founding Research Scientist at Cursive, working on large-scale post-training of long-horizon agentic models. My current work includes asynchronous reinforcement learning (async RL), on-policy self-distillation and memory-augmented models. I hold a PhD in Machine Learning from the University of Cambridge.

## Asynchronous RL and long-horizon training

At Cursive, I work on large-scale reinforcement learning and post-training for long-horizon agents. My notes explore the practical questions behind these systems: rollout throughput, policy lag, cheap proxy experiments and communication between agents. This work connects training infrastructure with the behavior of models over extended tasks.

https://www.pauliusrauba.com/research#async-rl

## AI safety, alignment and evaluation

How can we detect when a model fails, test whether its behavior changes, and control what it is allowed to do? My work includes context-aware testing, statistical audits of language-model robustness, least-privilege language models and generalization under distribution shift. My work on principal-agent problems examines questions of incentives and alignment in multi-agent systems.

https://www.pauliusrauba.com/research#safety-evaluation

## Efficient language models and adaptive computation

My research on deep hierarchical learning asks how a single model can support different computational budgets. Related work explores nested subspace networks, recursive models and masked diffusion. These papers sit alongside my current interests in memory-augmented models and efficient model deployment.

https://www.pauliusrauba.com/research#efficient-models

## Probabilistic forecasting and AI for science

At Salient, I worked on generative models for global weather forecasting. This research includes probabilistic transformers that jointly model atmospheric dynamics and decision-centric variables, and timestep-conditioned transformers for global forecasts. The papers link model design to the quantities that forecast users need.

https://www.pauliusrauba.com/research#weather-forecasting

## Papers and preprints

### Timestep-Conditioned Transformers for Global Weather Forecasting

Authors: Levang S., Bartolic F., Dickinson T., Dwelle C., Rauba P., Cikojevic V.

Year: 2026

Venue / status: arXiv preprint · 2608.06241

Source: https://arxiv.org/abs/2608.06241

### Demonstrating Generalization Failures via Mixtures of Conditional Policies

Authors: Barzdukas J., Peck J., Schulz J., Rauba P., Basart S., Wells L.

Year: 2026

Venue / status: arXiv preprint · 2607.03478

Source: https://arxiv.org/abs/2607.03478

### Recursive Scaling in Masked Diffusion Models

Authors: Carballo-Castro A., Piskorz J., Rauba P., van der Schaar M., Frossard P.

Year: 2026

Venue / status: arXiv preprint · accepted at BeNTo Workshop, NeurIPS 2026

Source: https://arxiv.org/abs/2606.18022

### Scaling with Recursion in Masked Discrete Diffusion Models

Authors: Carballo-Castro A., Piskorz J., Rauba P., van der Schaar M., Frossard P.

Year: 2026

Venue / status: FoGen Workshop @ ICML 2026

Source: https://openreview.net/pdf/3e382cf84f290264ffe4be6b1d375b0e261e8d2e.pdf

### Deep Hierarchical Learning with Nested Subspace Networks for Large Language Models

Authors: Rauba P., van der Schaar M.

Year: 2026

Venue / status: ICLR 2026

Summary: Nested Subspace Networks (NSNs) reparameterize each linear layer of a neural network as a low-rank product W = B A whose columns are nested, so that every rank-r sub-model is a strict prefix of every higher-rank one. A single set of weights then encodes a continuous hierarchy of models that can be traversed at inference time by simply choosing the rank, trading compute for accuracy with no retraining and no model zoo. Applied to large language models, NSNs unlock anytime inference, hardware-aware deployment, and elastic capacity under a fixed memory budget.

Source: https://openreview.net/forum?id=ymUOPsbxLi

### Multi-Agent Systems Should be Treated as Principal-Agent Problems

Authors: Rauba P., Cepenas S., van der Schaar M.

Year: 2026

Venue / status: arXiv:2601.23211

Source: https://arxiv.org/abs/2601.23211

### No More, No Less: Least-Privilege Language Models

Authors: Rauba P., Seputis D., Vanagas P., van der Schaar M.

Year: 2026

Venue / status: ICML 2026

Source: https://proceedings.mlr.press/v306/rauba26a.html

Source: https://arxiv.org/abs/2601.23157

### Probabilistic Transformers for Joint Modeling of Global Weather Dynamics and Decision-Centric Variables

Authors: Rauba P., Cikojevic V., Bartolic F., Levang S., Dickinson T., Dwelle C.

Year: 2026

Venue / status: Salient research model report

Source: https://arxiv.org/abs/2601.03753

### Tiny Autoregressive Recursive Models

Authors: Rauba P., Fanconi C., van der Schaar M.

Year: 2026

Venue / status: ICLR 2026 Workshop on AI with Recursive Self-Improvement · Spotlight

Source: https://openreview.net/forum?id=aY5kmaNrwB

### Model Organisms for Generalization Resistance Under Distribution Shift

Authors: Barzdukas J., Peck J., Schulz J., Rauba P., Wells L.

Year: 2026

Venue / status: ICLR 2026 Workshop on Principled Design for Trustworthy AI

Source: https://iclr.cc/virtual/2026/10019222

### Statistical Hypothesis Testing for Auditing Robustness in Language Models

Authors: Rauba P., Wei Q., van der Schaar M.

Year: 2025

Venue / status: ICML 2025

Summary: Distribution-Based Perturbation Analysis (DBPA) is a frequentist hypothesis-testing framework for auditing the robustness of large language models. Because LLMs are stochastic, comparing a single response under an original prompt to a single response under a perturbed prompt cannot reliably tell us whether the perturbation actually changed model behaviour. DBPA samples N responses under each prompt, embeds each response into a low-dimensional semantic space, and runs a permutation test on the centroid distance — yielding a p-value with no distributional assumptions. The framework supports arbitrary prompt perturbations and applies Bonferroni correction across multiple audits.

Source: https://openreview.net/forum?id=ECayXPDoha

### Visualizing Token Importance for Black-Box Language Models

Authors: Rauba P., Wei Q., van der Schaar M.

Year: 2025

Venue / status: AISTATS 2025

Source: https://proceedings.mlr.press/v258/rauba25a.html

### Redefining Digital Health Interfaces with Large Language Models

Authors: Imrie F., Rauba P., van der Schaar M.

Year: 2025

Venue / status: Frontiers in Artificial Intelligence

Source: https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2025.1623339/full

### Context-Aware Testing: A New Paradigm for Model Testing with Large Language Models

Authors: Rauba P., Seedat N., Luyten M.R., van der Schaar M.

Year: 2024

Venue / status: NeurIPS 2024

Summary: Context-Aware Testing (CAT) is a new paradigm for ML model testing that uses large language models to read the surrounding context of a task and generate the failure modes that actually matter, rather than exhaustively scanning data slices. The accompanying SMART method generates context-grounded hypotheses about model failures, refutes them with the data using statistical tests, and applies multiple-testing correction so that the set of surfaced failures has controlled false-discovery rate. CAT addresses a fundamental limitation of data-only testing — that exhaustive slice search inflates Type-I error and surfaces failures that are not domain-meaningful.

Source: https://openreview.net/forum?id=d75qCZb7TX

### Self-Healing Machine Learning: A Framework for Autonomous Adaptation in Real-World Environments

Authors: Rauba P., Seedat N., Kacprzyk K., van der Schaar M.

Year: 2024

Venue / status: NeurIPS 2024

Summary: Self-Healing Machine Learning (SHML) is a framework for models that autonomously adapt to distribution shift without human intervention. Existing concept-drift methods are reason-agnostic — they pick from a fixed menu of corrective actions without ever asking why the model is failing. SHML instead frames adaptation as an optimization problem over actions and uses an agentic system, ℋ-LLM, in which a language model diagnoses the structure of the shift (covariate, label, concept, or spurious-feature) and prescribes a corrective action that targets the actual cause. The framework provides theoretical guarantees on adaptation quality and empirically outperforms reason-agnostic baselines under real-world distribution shifts.

Source: https://openreview.net/forum?id=f63DKIpx0I

### Quantifying Perturbation Impacts for Large Language Models

Authors: Rauba P., Wei Q., van der Schaar M.

Year: 2024

Venue / status: Statistical Foundations of LLMs Workshop @ NeurIPS 2024

Source: https://arxiv.org/abs/2412.00868

### Armed Conflict and State Policy Preferences: A Statistical and Neural Embedding Approach

Authors: Rauba P.

Year: 2021

Venue / status: University of Oxford

Source: https://www.researchgate.net/profile/Paulius-Rauba-2/publication/356580919_Armed_Conflict_and_State_Policy_Preferences_A_Statistical_and_Neural_Embedding_Approach/links/61a21d097323543e210d5a4f/Armed-Conflict-and-State-Policy-Preferences-A-Statistical-and-Neural-Embedding-Approach.pdf

### Memes in the Wild: Assessing the Generalizability of the Hateful Memes Challenge Dataset

Authors: Kirk H.R., Jun Y., Rauba P., et al.

Year: 2021

Venue / status: WOAH @ ACL 2021

Source: https://aclanthology.org/2021.woah-1.4/

## Published technical notes

- [#4 The exact learning recipe is not always so important in long-horizon async RL](https://www.pauliusrauba.com/notes/4-the-exact-learning-recipe-is-not-always-so-important-in-long-horizon-async-rl): It's tempting to think that nailing the exact learning recipe is very important for online RL. While true in principle, there is a space of recipes that all give very…

- [#3 Notes on system performance of large-scale agentic models](https://www.pauliusrauba.com/notes/3-notes-on-system-performance-of-large-scale-agentic-models): In online RL, we often care about throughput, i.e. the number of rollouts that finish per hour. Let's understand the anatomy of throughput and how it works in agentic…

- [#2 Agent-to-agent communication via github](https://www.pauliusrauba.com/notes/2-agent-to-agent-communication-via-github): If you're collaborating with multiple people, you often need to exchange information that your agents have that you do not have. The fact that agents do not communicate…

- [#1 Principles for setting up a cheap proxy model for reinforcement learning training](https://www.pauliusrauba.com/notes/1-principles-for-setting-up-a-cheap-proxy-model-for-reinforcement-learning-training): How to choose and validate a cheap proxy model for reinforcement learning: configurations, LoRA, policy lag, and experiments with Kimi K3.1 and Qwen.

## Contact

For research conversations, collaborations or opportunities: https://www.linkedin.com/in/paulius-rauba/
