Publications

Discovery of Hidden Miscalibration Regimes

Kasia Kobalczyk, Mihaela van der Schaar

May 2026

Definition and discovery of hidden miscalibration regimes, providing insights into calibration structure and reliability of LLMs.

• model calibration
• uncertainty quantification
• probabilistic machine learning
• LLMs

LILO: Bayesian Optimization with Natural Language Feedback

Kasia Kobalczyk, Zhiyuan Jerry Lin, Benjamin Letham, Zhuokai Zhao, Maximilian Balandat, Eytan Bakshy

May 2026

A language-in-the-loop framework that leverages LLMs to transform free-form textual feedback into a proxy utility function, enabling principled Bayesian optimization with Gaussian Processes.

• Bayesian optimization
• decision making under uncertainty
• human-AI interaction
• Gaussian processes

Eliciting Numerical Predictive Distributions of LLMs without Autoregression

Julianna Piskorz^*, Kasia Kobalczyk^*, Mihaela van der Schaar

∗ Equal contribution

January 2026

Simple probes that recover the predictive distributions of LLMs including uncertainty estimates, sidestepping autoregressive sampling.

• uncertainty quantification
• probabilistic machine learning
• mechanistic interpretability
• LLMs

The Synergy of LLMs & RL Unlocks Offline Learning of Generalizable Language-Conditioned Policies with Low-fidelity Data

Thomas Pouplin^*, Kasia Kobalczyk^*, Hao Sun, Mihaela van der Schaar

∗ Equal contribution

May 2025

Training LLM agents as language-conditioned policies without expensive labels or online interaction, generalizing to unseen goals and states.

• reinforcement learning
• offline learning
• language-conditioned policies
• LLMs

Preference Learning for AI Alignment: a Causal Perspective

Kasia Kobalczyk, Mihaela van der Schaar

May 2025

A causal framework for preference learning that addresses causal misidentification, preference heterogeneity, and confounding from user-specific objectives.

• preference learning
• AI alignment
• causal inference
• human-AI interaction

Active Task Disambiguation with LLMs

Kasia Kobalczyk^*, Nicolás Astorga^*, Tennison Liu, Mihaela van der Schaar

∗ Equal contribution

January 2025

Formalises task ambiguity in natural-language specifications and frames disambiguation as Bayesian Experimental Design, leading to better clarifying questions from LLMs.

• Bayesian experimental design
• decision making under uncertainty
• human-AI interaction
• LLMs

Few-shot Steerable Alignment: Adapting Rewards and LLM Policies with Neural Processes

Kasia Kobalczyk, Claudio Fanconi, Mihaela van der Schaar

December 2024

Adapting reward models and LLM policies to user-specific preferences with few-shot learning, using neural processes to model reward uncertainty and heterogeneity.

• AI alignment
• reinforcement learning
• personalisation
• neural processes

Towards Automated Knowledge Integration From Human-Interpretable Representations

Kasia Kobalczyk, Mihaela van der Schaar

June 2024

Introduces informed meta-learning: an approach to inductive bias specification grounded in human knowledge, including unstructured natural language.

• meta-learning
• prior knowledge
• inductive bias
• probabilistic machine learning