1

Comparing how language models and humans balance compact representations with semantic detail.

Improving the reliability of fine-tuned foundation models with priors that account for uncertainty.

Adjusting dropout during inference to estimate uncertainty without additional labels or training.

Regularizing intermediate representations to reduce collapse and support sequential reasoning.

A decoding method that scales its token cutoff with the model's confidence, giving a simple way to control sampling as temperature changes.

An evolving language-model benchmark with recent questions and objective scoring, designed to limit test contamination.

We present OpenDebateEvidence, a massive dataset containing 3.5 million documents from competitive debate, enabling advancement in argument mining and summarization through large language model fine-tuning.

We provide an information-theoretic analysis of VICReg, deriving theoretical foundations for deterministic networks and introducing new SSL methods based on these insights.

We show that carefully tuning standard deep learning components can achieve state-of-the-art performance on class-imbalanced datasets without specialized techniques.

We examine self-supervised learning methods to provide an information-theoretical understanding of their construction. As a first step, we demonstrate how information-theoretic quantities can be obtained for a deterministic network. This enables us to demonstrate how SSL methods can be (re)discovered from first principles and their assumptions about the data distribution. Furthermore, we empirically demonstrate the validity of our assumptions, confirming our novel understanding.